Professional Documents
Culture Documents
Closed Form Approximation of LoRa Modulation BER Performance
Closed Form Approximation of LoRa Modulation BER Performance
Closed Form Approximation of LoRa Modulation BER Performance
fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LCOMM.2018.2849718, IEEE
Communications Letters
Abstract—LoRa has been demonstrating itself as a front then derived in Section III and their accuracy is evaluated in
runner when it comes to evolving Low Power Wide Area Section IV. Finally, Section V is dedicated for conclusions.
Networks (LP-WAN). At the core of LoRa is LoRa modulation
which depicts a patented chirp spread spectrum modulation II. BASICS OF L O R A C HIRP M ODULATION
technique that supports energy-efficient and reliable long-range
communication. The underlying BER performance of LoRa LoRa adopts a frequency shift chirp modulation scheme
modulation has not been yet rigorously analyzed in the liter- [2]. Consider a LoRa system in the baseband where the chirp
ature. In this letter, closed form approximations for LoRa BER signals have bandwidth B . Consequently, one LoRa sample is
performance are derived for both AWGN and Rayleigh fading transmitted every T = 1/B . LoRa modulation is achieved by
channels. Accuracy of the derived approximations is confirmed
by comparisons against numerical results. spreading the frequency change of the chirp signal across 2SF
Index Terms—LP-WAN, LoRa, Chirp Modulation, Spread samples within a symbol duration of TS = 2SF ·T , where SF ∈
Spectrum, BER Performance Analysis. {7, 8, ..., 12} depicts the spreading factor. The LoRa encoder
I. I NTRODUCTION maps every SF bits to a symbol sk = k, k ∈ {0, 1, 2, . . . , 2SF −
1}. Each LoRa symbol sk is characterized by a frequency
Common consensus for the realization of a multitude of offset fk = B · k/2SF that sets the starting frequency of the
Internet-of-Things (IoT) scenarios particularly within smart corresponding chirp signal. During the symbol duration TS ,
cities envisions deployment over Low Power Wide Area Net- the signal frequency corresponding to sk increases linearly
works (LP-WAN) [1]. The terminology reflects a star topology with a slope of B 2 /2SF from fk until it reaches B . The chirp
where thousands of power-constrained IoT devices communi- signal frequency then wraps around to 0 before it continues to
cate directly with an IoT gateway over long distances at the increase linearly again until the end of TS . With a sampling
scale of kilometers [1]. Recently, LoRa has been exhibiting rate of 1/T , the LoRa waveforms wk (nT ), k ∈ {0, 1, 2.., 2SF −
tremendous commercial growth to establish itself among the 1}, corresponding to the LoRa symbols sk are comprised of
front runners of emerging LP-WAN. LoRa has been primarily 2SF samples and are expressed as
pillared on the patented LoRa modulation. LoRa modulation √
features a chirp spread spectrum modulation that supports wk (nT ) = ES × ωk (nT ) =
r h
energy-efficient/reliable long-range communication. Given its ES n i
patented protection, thorough theoretical description for LoRa SF
exp j2π · (k + n) mod 2SF · , (1)
2 2SF
modulation may still be considered sparse in the scientific
research community state-of-the-art. In fact, the first research where ωk (nT ), k ∈ {0, 1, 2..., 2SF −1}, are the 2SF orthonormal
to shed some light on the mathematical representation and basis functions that characterize the multi-dimensional space
signal processing-based implementation for LoRa modulation of LoRa signaling, ES depicts the symbol energy, and n =
has been recently presented in [2]. The author in [2] also 0, 1, 2, . . . , 2SF − 1 depicts the sample index at time nT .
refers to some high level description presented in [3] as well LoRa demodulation is based on the orthogonality of chirp
as some theoretical analog-based description in [4], [5]. While signals with different offsets [2], where the correlation prop-
the theoretical description of LoRa modulation is becoming erties of the 2SF possible LoRa waveforms are governed by
2SF
(√
evident, to the best of the authors’ knowledge, the theoretical X−1 ES i=k
BER analysis of LoRa modulation is still completely lacking. wk (nT ) · ωi∗ (nT ) = , (2)
0 i 6= k
The only formula for LoRa BER performance reported in n=0
the literature has been estimated through curve fitting of where ωi∗ (nT ) is the complex conjugate of the basis function
simulated data in [5]. Nevertheless, LoRa modulation may still ωi (nT ) that corresponds to the LoRa symbol si = i.
be categorized under the classical problem of error probability Accordingly, for a received signal rk (nT ) that corresponds
of orthogonal signaling with non-coherent detection. The to a transmitted waveform wk (nT ) over an AWGN channel,
underlying BER performance has been well-established to be the correlator outputs of the LoRa demodulator are given as
evaluated numerically rather than in closed form [6] (sec. 4.5- 2SF
(√
X−1 E S + φi i=k
3 p. 216). Accordingly, the main contribution of this letter is rk (nT ) · ωi∗ (nT ) = , (3)
the derivation of closed form approximations for LoRa BER n=0
φi i 6= k
performance in both AWGN and Rayleigh fading channel
where φi depicts a complex Gaussian noise process.
environments. The approximations rely on the fact that LoRa
The detection of LoRa symbols reverts to selection of the
features a relatively high order of M -ary modulation with
index of the LoRa waveform that has the highest correlation
M ≥ 128. The accuracy of the derived approximations is then
magnitude with the received signal. Therefore, for a received
confirmed by comparisons against numerical results.
waveform rk (nT ) that corresponds to a transmitted symbol
The rest of this letter is organized as follows. In Section II,
sk = k, the detected symbol sek is computed as
theoretical description of LoRa modulation is presented. The √
closed form approximations for LoRa BER performance are sek = argi max δk,i ES + φi , (4)
1089-7798 (c) 2018 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.2018.2849718, IEEE
Communications Letters
Es /TS ES
where δk,i = 1 for i = k and δk,i = 0 otherwise. Γ= = , (10)
N0 · B N0 · 2SF
Let us define ρi = |φi | as the magnitude of the com-
plex noise envelope. Given that φi is a complex zero-mean then the shape parameter for the Rician distributed β can
Gaussian noise process, then ρi depicts a Rayleigh distributed be expressed as κβ = Γ · 2SF . In other words, due to
random variable with the cumulative distribution function spreading in LoRa, the target symbol enjoys an effective SNR
of Γef f = Γ · 2SF . The Rician distribution for high SNR can
ρ2
Fρi (ρ) = 1 − exp − 2 , (5) be approximated with a Gaussian distribution [7]. Since LoRa
2σ
modulation deploys SF ≥ 7, the scale of attainable Γef f may
where σ 2 = N0 /2 and N0 is the single-sided noise power justify a Gaussian approximation for fβ (β). Hence,
spectral density. " √ 2 #
1 β − ES
Let us define Pe|k as the probability of symbol error given fβ (β) = √ exp − , (11)
2πσ 2 2σ 2
that symbol sk = k is transmitted. Given the outputs of the
correlation process defined in (3), Pe|k can be expressed as √
where ES and σ 2 = N0 /2 are the mean and variance of β ,
respectively.
Pe|k = Pr max(ρi ) > βk , (6)
i,i6=k Let us also assume that the distribution of ρ̂ for the
√ maximum of 2SF − 1 i.i.d. Rayleigh random variables follows
where βk = ES + φk and βk accordingly follows a Rician
a Gaussian distribution. Then, we get
distribution with the shape parameter κβ = ES /2σ 2 = ES /N0 . " #
Let ρ̂ = maxi,i6=k (ρi ) depict a random variable for the 1 (ρ − µρ̂ )2
fρ̂ (ρ) = q exp − , (12)
maximum of 2SF − 1 i.i.d. Rayleigh random variables. The 2πσρ̂2 2σρ̂2
cumulative distribution function for ρ̂ can be given as
2SF −1 where µρ̂ and σρ̂2 are the mean and variance for ρ̂, respectively.
ρ2
Fρ̂ (ρ) = 1 − exp − 2 . (7) From (6) and assuming that every symbol error entails an
2σ
underlying 50% of bits errors, the bit error probability Pb is
Given (6), (7) and assuming equally probable symbols, the expressed as
average bit error probability Pb can be expressed as Pb = 0.5 × Pr [ρ̂ − β > 0] . (13)
Z∞ " 2SF −1 #
It follows from the Gaussian approximations for ρ̂ and β that
β2
Pb = 0.5 × 1 − 1 − exp − 2 · fβ (β)dβ, (8) the bit error probability then becomes
2σ
0 √
E S − µρ̂
where fβ (β) is the probability density function for the Rician Pb ≈ 0.5 × Q q , (14)
distributed β . Given that LoRa falls under the category of σρ̂2 + N0 /2
orthogonal signaling, the multiplier of 0.5 before the integral R h 2i
reflects the fact that for each symbol error only half of the with Q (z) = √12π . z∞ exp − u2 du being the Q-function, i.e.,
underlying symbol bits are expected to be in error [6] (sec. the tail function of the standard normal distribution. It is
4.4-1 p. 205). The integral in (8) could be shown to be evident that a closed form approximation for Pb in (14) is
evaluated as [6] (sec. 4.5-3 p. 218) attainable if µρ̂ and σρ̂2 corresponding to the mean and variance
! of ρ̂ = maxi,i6=k (ρi ), respectively, are derived. Let us define
2SF
X−1 (−1)q+1 2SF −1 q ES γi = ρ2i , where ρi is Rayleigh distributed and governed by
Pb = 0.5 × . . exp − . . (9)
q=1
q+1 q q + 1 N0 (5). The distribution of γi is Chi-square with one degree of
freedom, where γi has the cumulative distribution function
Given that LoRa depicts a relatively high order of M - h γ i
ary modulation that reaches M = 4096 when SF = 12, the Fγi (γ) = 1 − exp − 2 . (15)
2σ
combination term in (9) would suffer from precision problems Let γ̂ = maxi (γi ) depict the maximum of 2SF − 1 variables
during the evaluation of Pb . The contribution of this letter is √
γi . Accordingly, ρ̂ = γ̂ and the following relations hold
that it uses the high order feature of LoRa modulation to derive
an approximation for the distribution of ρ̂. This in turn leads E ρ̂2 = E [γ̂] = µγ̂
to the derivation of concise closed form and highly accurate E ρ̂4 = E γ̂ 2 = σγ̂2 + µ2γ̂ , (16)
approximation expressions for (8).
where E [X m ] is the mth moment for the distribution of
III. C LOSED F ORM A PPROXIMATION FOR L O R A some variable X and µγ̂ , σγ̂2 are the mean, variance of γ̂ ,
M ODULATION BER P ERFORMANCE respectively. Assuming a Gaussian distribution for ρ̂, E [ρ̂m ]
A. LoRa BER Performance in AWGN Channels could be derived from the mth differential
2 2
of the moment
generating function Mρ̂ [t] = e(µρ̂ t+σρ̂ t /2) evaluated at t = 0.
As presented in (8), the BER performance of LoRa mod- It is then straightforward to derive the following relations
ulation is attributed to the probabilistic comparison of a
Rician distributed variable against the maximum of 2SF − 1 E ρ̂2 = σρ̂2 + µ2ρ̂
Rayleigh distributed variables. Let us first focus on the Rician E ρ̂4 = 3σρ̂4 + 6σρ̂2 µ2ρ̂ + µ4ρ̂ . (17)
distribution with shape parameter κβ = ES /N0 that reflects the
correlator output when correlating with the basis ωk (nT ) that By combining (16) and (17) it is possible to show that
1/4
corresponds to the actual transmitted symbol sk = k. Let us µρ̂ = µ2γ̂ − σγ̂2 /2
define Γ as the SNR for the LoRa communication. Given that q
σρ̂2 = µγ̂ − µ2γ̂ − σγ̂2 /2. (18)
1089-7798 (c) 2018 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.2018.2849718, IEEE
Communications Letters
For large m, the harmonic number Hm approaches Using Leibniz’s integral rule, the differentiation term in (29)
1 could be given as
Hm ≈ ln(m) + + 0.57722, (22)
2m d h p p i
where 0.57722 is the Gamma or the Euler-Mascheroniconstant Q 2αΓef f − 2H2SF −1 =
2 2
dα √
[9]. Given that SF ≥ 7, we can assume that H2SF −1 ≫ π12 . √2H 2
1 −
2αΓef f −
2SF −1 p 1
α− 2
Moreover, if we consider that ln 2SF − 1 ≈ ln 2SF , a more −√ e 2 · 2Γef f . (30)
concise approximation for Pb becomes 2π 2
√ √ Using a change of variables and substituting for (30) in
Pb ≈ 0.5 × Q Γ · 2SF +1 − 1.386 · SF + 1.154 . (23) (29), the approximation for probability of bit error becomes
It worth noting that the approximation in (23) is consistent p 1
Pb ≈ 0.5 × Q − 2H2SF −1 − √
with the form provided in [5] that has been attained using 2π
curve fitting of simulation results and has been expressed as
√
Z∞ (Γef f +1)y2 +2 2H2SF −1 y+2H2SF −1
−
√ √ × e 2Γef f
dy . (31)
Pb = 0.5 × Q 1.28 · Γ · 2SF − 1.28 · SF + 0.4 . (24) √
− 2H2SF −1
B. LoRa BER Performance in Rayleigh Channels
By completing the square of the quadratic equation in the
In the case of Rayleigh channels, the correlation outputs at exponent and another substitution of variables, (31) becomes
the LoRa demodulator from (3) are alternatively given as
p
Pb ≈ 0.5 × Q − 2H2SF −1
2SF
(√
X−1 αES + φi i=k s
rk (nT ) · ωi∗ (nT ) = , (25) −
2H SF
2 −1 Z∞
Γef f 1 − u22
n=0
φi i 6= k − .e 2(Γef f +1) × √ e du , (32)
√ Γef f + 1 2π
where α is the Rayleigh distributed complex envelope am- x
q √
plitude and α is the corresponding instantaneous normalized Γef f +1 p 2H2SF −1
where x = − 2H2SF −1 + . Given that
channel power. Accordingly, for the Rayleigh fading scenario, Γef f Γef f +1
the bit error probability in (13) could be expressed as the integral in (32) depicts the Q-function and Γef f = 2SF Γ,
1089-7798 (c) 2018 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.2018.2849718, IEEE
Communications Letters
10 0
the approximation for LoRa probability of bit error within a
Rayleigh fading environment finally reduces to
10 -1
r 2SF Γ
2H SF
2
p
Figure 1 presents comparisons of the derived LoRa BER SF = 12, Approx. (21)
SF = 12, Approx. (23)
SF = 12, Approx. (24) [5]
10 -6
approximations in (21) and (23) against the theoretical BER -25 -24 -23 -22 -21 -20 -19 -18 -17 -16 -15 -14
in (8) at SF = 10, 11, 12. The figures also evaluate the Fig. 1. LoRa BER Performance in AWGN Channel: SF = 10, 11, 12.
curve fitting solution (24) from [5]. As shown in the fig- SF=7 Rayleigh Channel Numerical
ure, the expression in (21) exhibits an extremely accurate SF=7 Rayleigh Channel Approx. (33)
SF=7 AWGN Channel Approx. (23)
1089-7798 (c) 2018 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.