IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 21, NO.

1, JANUARY 2022 357

On the LoRa Chirp Spread Spectrum Modulation:

Signal Properties and Their Impact on Transmitter
and Receiver Architectures
Gianni Pasolini , Member, IEEE

Abstract— The LoRa modulation scheme is arousing a growing that more than 50% of all non-cellular LPWAN connections
interest in the Internet of Things community as it is adopted by will be LoRa-enabled by 2026, this technology is predicted to
the emerging LoRaWAN technology. In this paper, we firstly dominate the LPWANs market in the next years.
analyse the baseband processing for the generation of LoRa
signals at the transmitter side, providing a simple algorithm From a technical point of view, LoRaWANs are defined
that leverages digital signal processing techniques to reduce the by an open standard [10] developed by the LoRa Alliance,
modulator complexity. Secondly, we analytically investigate the which mainly defines the Medium Access Control (MAC)
signal demodulation technique. Quite surprisingly, we found that layer and the message formats. At the physical (PHY) layer,
its effectiveness depends on the particular choice of the sampling they are based on LoRa, a proprietary modulation scheme
frequency at the receiver side, which purposely does not meet
the sampling theorem requirement. Finally, we consider the developed by Semtech Corporation and derived from Chirp
actual architecture of digital receivers investigating the trade-off Spread-Spectrum (CSS).
between the selectivity of receive digital filters, which impacts on In the following sections, we provide an in-depth analysis of
the required computational effort and power consumption, and the LoRa PHY layer. We start by investigating the generation
the receiver performance. of the modulated signal at the transmitter, which is addressed
Index Terms— Digital modulation, Chirp Spread-Spectrum leveraging digital signal processing techniques. We also focus
(CSS), LoRa, Low-Power Wide-Area Networks (LPWANs), the attention on the receiver, providing a frequency-domain
Internet of Things (IoT). analysis of the demodulation technique. Although the numer-
ical demodulation algorithm adopted by LoRa receivers is
I. I NTRODUCTION
well-known, since it is described in [11], [12] and discussed

T HE term Internet of Things (IoT) was first used in 1999 to

describe a scenario where the Internet is connected to the
physical world via ubiquitous sensors [1]. Two decades on,
in several excellent papers, such as [13], [14], no paper has
dealt so far with the frequency-domain implications of each
of its steps, which however reveal the brilliant idea behind its
a lot of progress has been made in this direction, as several remarkable performance.
communication technologies, both short- and long-range, have Finally, we also investigate the receiver performance as a
been developed, which are capable of supporting IoT applica- function of the signal-to-noise ratio. This topic is covered
tions in different contexts and environments. by many papers, such as [15]–[19]. However, all of them
Focusing the attention on long-range technologies, assume ideal filtering at the receiver side, which may be
Low-Power Wide-Area Networks (LPWANs) have recently impossible to implement, due to constraints on the available
appeared on the IoT stage [2], the most popular being computational power or the need to reduce the energy con-
LoRaWAN, SigFox, and NB-IoT [3]. Such networks are pecu- sumption. In this paper, we consider the real architecture of
liar in that they combine wide coverage (even several kms) modern receivers, which usually include digital filters, and we
and long battery life (even ten years), at the cost of a low bit investigate the impact of the filters’ selectivity on the receiver
rate. performance.
At the time of writing, LoRaWAN represents the The main contributions of this paper can be summarized as
“de-facto” standard for the Internet of Things, being installed follows:
in 163 countries by 150 network operators [4] to enable a • we provide an original strategy for the generation of
wide range of services [5]–[8]. With more than 170 million LoRa signals, which leverages digital signal processing
of already connected devices [9], and with the expectation techniques to reduce the transmitter complexity (Sec.IV),
• we analytically investigate the demodulation technique
Manuscript received November 16, 2020; revised April 26, 2021; accepted
June 22, 2021. Date of publication July 15, 2021; date of current version described in the patent documents [11], [12] filed by
January 10, 2022. The associate editor coordinating the review of this article Semtech, revealing still unknown aspects, mainly con-
and approving it for publication was X. Tao. cerning the impact of signal undersampling, that are the
The author is with the Department of Electrical, Electronic and Information
Engineering “G. Marconi,” University of Bologna, 40136 Bologna, Italy, and key elements for its effectiveness (Sec.V),
also with the National Laboratory of Wireless Communications (WiLab) of • we investigate the trade-off between the receiver perfor-
CNIT, 40136 Bologna, Italy (e-mail: [email protected]). mance and the complexity of digital receive filters, which
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TWC.2021.3095667. impacts on the energy consumption and the required
Digital Object Identifier 10.1109/TWC.2021.3095667 computational effort (Sec.VI).
1536-1276 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: INP TOULOUSE. Downloaded on June 22,2022 at 12:56:27 UTC from IEEE Xplore. Restrictions apply.
358 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 21, NO. 1, JANUARY 2022

TABLE I
TABLE OF S YMBOLS AND R ELATED M EANINGS

Notation: Throughout this paper, we define the indicator

function gT (t) = 1 for 0 ≤ t < T and gT (t) = 0 Fig. 1. Instantaneous frequency-offset. s = 91, SF = 8, B = 500 kHz,
elsewhere and we denote by j the imaginary unit. For the SF
Ts = 2B = 5.12 · 10−4 s.
reader’s convenience, in Table I we reported the main symbols
introduced in the following, along with their meanings. signal modulated by random data is investigated, obtaining a
closed-form expression of the spectrum in terms of Fresnel
II. O RIGINAL C ONTRIBUTIONS W ITH R ESPECT functions. Although very thorough and inspiring, the analyses
TO THE E XISTING L ITERATURE
provided in [13], [14] do not address the fundamental role
LoRa signals have been investigated in several papers. played by undersampling and aliasing in the demodulation
Those that are most related to the content of this article are process, which is instead the original contribution reported in
discussed below. Sec.V of this paper.
The generation of LoRa modulated chirps starting from Regarding the performance of the LoRa modulation, many
the time-shift of a basic chirp was proposed in 2019 by papers present analytical or simulative investigations. In [15]
Nguyen et.al. [20] and Elshabrawy and Robert [21], which and [16], for instance, the symbol error rate (that is, the chirp
are rightly credited with being the first to introduce numerical error rate) is derived as a function of the signal-to-noise ratio.
processing techniques to reduce the complexity of LoRa trans- In [17], the closed-form expression of the bit error proba-
mitters. However, the signal generated adopting the techniques bility is derived in AWGN and Rayleigh channels. Similar
presented in [20], [21] does not exhibit the continuous-phase results are presented in [18], [19], [23] for different signal
property when passing from one symbol (i.e., one chirp) to parameters, channel conditions and demodulation techniques.
the following one, which was demonstrated in the same year However, none of the papers in the literature consider the
by Chiani and Elzanaty [14]. The enhancement proposed by trade-off between the performance and the complexity of the
Elshabrawy et.al. in [22], while introducing phase continuity, digital filters implemented in LoRa receivers, which is instead
generates a discrete-time signal with a sampling rate that addressed in Sec.VI of this paper.
might be critical for real implementations. This aspect will The analysis of the LoRa physical layer reported in the
be discussed in Sec.IV, where an original expression of the above referenced articles and in this manuscript would not
sampled phase of LoRa chirps is provided. be complete without mentioning the impact of interference
The characteristics of LoRa signals, both in the time and and capture effect on the receiver performance. These topics
frequency domains, have been instead investigated in [13] are out of the scope of this document, mainly because they
and [14], which also discuss the demodulation strategy. have been specifically addressed by many papers. In [24] the
In [13], Vangelista gives the first mathematical description interference between LoRa signals is analysed for different
of the modulation and demodulation processes, providing spreading factors and bandwidths, whereas in [25], [26] the
the theoretical derivation of the optimum receiver based on use of successive interference cancellation is considered to
the Fast Fourier Transform. The performance of the LoRa decode superposed LoRa signals. The capture effect is exper-
modulation is also derived both in an additive white Gaussian imentally investigated in [27], [28] and analytically investi-
noise (AWGN) channel and in a frequency selective channel. gated in [26], [29], [30]. Both analysis and simulations are
In [14], Chiani and Elzanaty analysed the LoRa M -ary mod- instead discussed in [15], [31].
ulation, providing the signal description in the time domain
and showing that LoRa is a memoryless continuous phase
III. L O R A S IGNAL
modulation. The cross-correlation between the transmitted
waveforms is also derived in [14], proving that LoRa can be LoRa is a proprietary spread-spectrum modulation scheme
considered approximately an orthogonal modulation only for patented by Semtech Corporation [32], which is derivative
large M . Finally, in [14] the spectral characteristics of the of CSS. In general, a sine-wave chirp signal, concisely denoted

Authorized licensed use limited to: INP TOULOUSE. Downloaded on June 22,2022 at 12:56:27 UTC from IEEE Xplore. Restrictions apply.
PASOLINI: ON Lo, vol. 21, no. 1, january 2022Ra CSS MODULATION: SIGNAL PROPERTIES AND THEIR IMPACT 359

Fig. 2. Instantaneous frequency-offset and phase of a pure upchirp. SF = 12, B = 500 kHz.

as chirp in the following, consists of a short-time frequency with Th = Ts − TMs s denoting the time instant in which
sweep mathematically expressed as Δf (s, t), having reached B2 , wraps around to − B2 .
In the specific case of LoRa, the modulation parameters are
  t 
chosen such that
c(t) = V0 cos 2π f (ξ)dξ + φ0 , t0 ≤ t ≤ t0 + Tc (1) • B ∈ {125, 250, 500} kHz,
t0 SF
• M = 2 , with SF denoting the spreading factor,
where • SF ∈ {7, 8, 9, 10, 11, 12},
• V0 > 0 is the chirp amplitude,
• BTs = M .
• f (t) is the instantaneous frequency,
In Fig.1, an example of instantaneous frequency-offset with
• φ0 is the signal phase at the initial instant t0 ,
s = 91, B = 500 kHz and SF = 8 is given: one observes that
• Tc is the chirp duration.
Δf (s = 91, t) linearly sweeps within the interval [− B2 , B2 ]
Denoting with f0 the central frequency of the sweep interval and wraps around at Th = 3.3 · 10−4 s.
[f0 − B2 , f0 + B2 ] and assuming, without loss of generality, According to [33], [34], a specific frame format is used at
t0 = 0 and φ0 = 0, (1) can also be written as the physical layer of Semtech devices to accommodate data:
  t 
A LoRa frame begins with a preamble, that consists of a
c(t) = V0 cos 2πf0 t + 2π Δf (ξ)dξ , 0 ≤ t ≤ Tc (2)
0 configurable number of pure upchirps, whose frequency-offset
with Δf (t) representing the instantaneous frequency-offset, linearly increases from − B2 to B2 (see Fig.2(a) for the
ranging in the interval [− B2 , B2 ], with respect to f0 . frequency-offset of a pure upchirp in the case SF = 12 and
Equation (2) is the basis of the LoRa modulation, which B = 500 kHz), followed by two and a quarter downchirps,
uses M differently shaped chirps, each of which is in one- and, optionally, a frame header. The remainder of the frame
to-one correspondence with the M symbols of the modu- includes the payload, which carries ns symbols, and, option-
lation alphabet S = {0, · · · , M − 1}. In particular, given a ally, the corresponding CRC.
modulation symbol s ∈ S, the instantaneous frequency-offset Focusing on the payload, one observes that it consists of
Δf (s, t) of the corresponding modulated chirp linearly a sequence of ns modulated chirps, each of which with an
instantaneous frequency-offset similar to the one depicted
increases starting from − B2 + M B
s. Then, when the maximum
B in Fig.1 (they differ each other by the “wraps around
frequency-offset 2 is reached, Δf (s, t) wraps around to
instant” Th , which depends on s). In this regard, one observes
− B2 and keeps on increasing linearly. Assuming t0 = 0
that, irrespectively of the modulation symbol s, it is1
as the symbol starting instant, the chirp stops when  Ts
Δf (s, t = Ts ) = Δf (s, 0). Its duration Ts is usually referred
Δf (s, ξ)dξ = 0. (4)
to as symbol time. 0
In mathematical terms, still considering the symbol interval This means that the phase term
[0, Ts ], for any given modulation symbol s it is:  t
⎧ θ(s, t) = 2π Δf (s, ξ)dξ (5)
⎪ B B B
⎨− + s + t, 0 ≤ t < Th , 0
Δf (s, t) = 2 M Ts (3)
⎪
⎩−
3B
+
B B
s + t, Th ≤ t ≤ Ts
1 The proof, based on geometric reasoning, is trivial and is omitted for the

2 M Ts sake of conciseness.

Authorized licensed use limited to: INP TOULOUSE. Downloaded on June 22,2022 at 12:56:27 UTC from IEEE Xplore. Restrictions apply.
360 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 21, NO. 1, JANUARY 2022

in (2) is such that

θ(s, Ts ) = θ(s, 0) = 0. (6)

The proof is reported in Appendix A.

Equation (6) immediately results in the following two
properties:
• Property 1: The LoRa signal exhibits continuous phase
when passing from one modulation symbol to the follow-
ing one. In particular, the phase at the beginning and at Fig. 3. Transmitter: Quadrature modulator architecture. The triangle labelled
the end of each symbol interval is equal to zero. “PA” represents the power amplifier.
• Property 2: In any given symbol interval, irrespectively
on the symbol s to be transmitted, the shape of the requires the generation of the in-phase and quadrature base-
modulated chirp does not depend on the previously band signals p(t) = {i(t)} and q(t) = {i(t)}:
transmitted symbols, as the initial and final phases in a
s −1
n  
symbol interval are coincident.
p(t) = V0 cos θ(s(n) , t) gTs (t − nTs ) (10)
The above properties, which were first derived in [14] n=0
(albeit in a different way), imply that LoRa is a memoryless n s −1  
continuous phase modulation. Properties 1 and 2 are here q(t) = V0 sin θ(s(n) , t) gTs (t − nTs ). (11)
recalled because they are particularly useful from a practical n=0
point of view, as they ease the generation of the modulated In the perspective of a digital implementation, this means
signal at the transmitter side. This aspect is discussed in the that the baseband stage of the transmitter needs to generate
following section. the discrete-time signals
s −1
n  
IV. L O R A S IGNAL G ENERATION p(kT ) = V0 cos θ(s(n) , kT ) gTs (kT − nTs ) (12)
n=0
Since the waveform transmitted in each symbol interval n s −1  
depends only on the symbol to be transmitted in that interval q(kT ) = V0 sin θ(s(n) , kT ) gTs (kT − nTs ), (13)
(Property 2), the generation of LoRa signals boils down to n=0
the generation of a sequence of independently modulated
with k ∈ Z and T denoting a properly chosen sampling
chirps, each of which starting and ending with the phase
interval. These signals are then converted from digital to
term equal to zero (Property 1). In particular, denoting with
analog in the final stage of the baseband modulator, which
Δf (s(n) , t) the instantaneous frequency-offset
related

to the
includes a pair of digital-to-analog converters (DACs)3 for the
symbol s(n) in the time interval nTs , (n + 1)Ts , and with
generation of (10) and (11).4
θ(s(n) , t) the corresponding phase term, the modulated chirp
Clearly, the sampling frequency fs = T1 must be such
to be transmitted in the same interval is
  that the Shannon-Hartley sampling theorem is fulfilled with
 t
reference to p(t) and q(t), whose generation technique is
cn (t) = V0 cos 2πf0 t + 2π Δf (s(n) , ξ)dξ
the final objective of the present section. In this regard,
 nTs
 it is worth stressing that having the instantaneous frequency
(n)
= V0 cos 2πf0 t + θ(s , t) (7) sweeping over B does not imply that the bandwidth of the
radiofrequency (RF) signal is B. In particular, Chiani and
with n = 0, 1, . . . , ns − 1, and nTs ≤ t ≤ (n + 1)Ts .
Elzanaty [14] (Table I) show that the largest bandwidth occu-
The corresponding complex envelope in the same symbol
pation (i.e., the worst case for the sampling theorem) occurs
interval is thus
when SF = 7. In such a case, it is B99 = 1.045B, with B99
(n)
in (t) = V0 ejθ(s ,t)
. (8) denoting the RF bandwidth centred on f0 containing 99% of
the power.
Given (8), due to the memoryless nature of the modulation, Since the discussion herein considers the baseband compo-
the complex envelope of the LoRa signal2 can be written as nents p(t) and q(t) of LoRa RF signals, one should assume a
bandwidth occupation of p(t) and q(t) equal to B299 . It follows
s −1
n
(n) that fs should be such that fs ≥ 2 B299 , that is, in the worst
i(t) = V0 ejθ(s ,t)
gTs (t − nTs ). (9) case, fs ≥ 1.045B. As shown hereafter, a convenient choice
n=0
is fs = 2B, which largely fulfills the sampling theorem
As shown also in the block diagram reported in the LoRa requirement.
patent application EP2763321A1 [11], transmitters usually 3 The DACs are not shown in Fig.3, as they are inside the Baseband
adopt a quadrature modulator architecture (see Fig.3), which Modulator block.
4 Alternatively, the sampled in-phase and quadrature waveforms might be
2 Here, we do not consider the preamble, which does not depend on multiplied by the sampled in-phase and quadrature sinusoids, summed up and
modulation symbols. then digital-to-analog converted. This solution is not reflected in Fig.3.

Authorized licensed use limited to: INP TOULOUSE. Downloaded on June 22,2022 at 12:56:27 UTC from IEEE Xplore. Restrictions apply.
PASOLINI: ON Lo, vol. 21, no. 1, january 2022Ra CSS MODULATION: SIGNAL PROPERTIES AND THEIR IMPACT 361

One might observe that B99  B, so that the bandwidth of In the following, we show that starting from the
p(t) and q(t) can be assumed equal to B2 without a significant Nmax samples of the upchirp phase (16) when SF = 12,
lack of accuracy. Although this is certainly true, it does not it is possible to derive the sampled phase terms for all
mean that choosing fs = B is a reasonable choice, as it would possible modulation symbols and all possible values of SF,
appear by simply considering the sampling theorem condition. with significant benefits in terms of transmitter complexity.
In fact, when it comes to actually implement the transmitter For this reason, hereinafter we will refer to
with digital processing techniques, one has to consider that any    213 −1
 213 −1 π k
DAC includes an analog low-pass filter, aimed at preserving (12)
θ (k) = k −1 + 13 , (17)
undistorted the spectrum in the frequency interval [− B2 B2 ] k=0 2 2
k=0
and removing the periodic spectral replicas that occur with
period fs . Choosing fs = B would mean that the spectral as reference upchirp phase. Its shape is depicted in Fig.2(b).
replicas are perfectly adjacent (actually, there might even be a
little overlap at their boundaries for low SFs), thus leaving no B. Step 2: Sampled Upchirp Phase Term for Each SF
room for the DAC’s filter to completely remove the spectral The sampled upchirp phase for all SFs (given by (16)) can
components in the immediate vicinity of the interval [− B2 B2 ]. be derived by properly decimating and scaling (17). More
The remainder of this section is aimed at showing how the precisely, having defined the decimation factor D = 212−SF ,
baseband modulator of a LoRa digital transmitter can generate it is:
the baseband discrete-time signals (12) and (13) with fs = 2B.  2SF+1 −1 1  (12) 2SF+1 −1
θ(SF) (k) = θ (Dk) , (18)
k=0 D k=0
A. Step 1: Generation of the Reference Phase Term The proof of (18) is reported in Appendix C.
One considers, firstly, a pure upchirp, whose instantaneous It follows that the 213 = 8192 samples of (17) might be
frequency-offset Δf (t) linearly increases from − B2 to B2 . stored in a lookup table (LUT), so that the sampled upchirp
Fig.2(a) shows an example of pure upchirp in the case phase for all SFs can be derived (according to (18)) by simply
1
B = 500 kHz and SF = 12. The analytical reading 1 sample every D in the LUT and scaling it by D .
 t expression of
the corresponding phase term θ(t) = 2π 0 Δf (ξ)dξ in the Alternatively, the sampled upchirp phase for a given SF might
symbol interval [0, Ts ] can be easily derived: be computed run-time using (18), one sample at a time in each
  sampling interval T , thus avoiding the need of a LUT.
B B 2 Equation (18), which refers to the sampled phase of pure
θ(t) = 2π − t + t , 0 ≤ t ≤ Ts . (14)
2 2Ts upchirps, is only an intermediate step toward our final objec-
Since (14) refers to a pure upchirp, in the following tive, which is the derivation of a similar equation expressing
it will be denoted as upchirp phase. One considers, now, the sampled phase for all modulation symbols s and all SFs.
the discrete-time upchirp phase Nonetheless, (18) may be directly used to generate the cos(·)
and sin(·) terms in (12) and (13) when pure upchirps need
 N −1   to be transmitted, that is, in the frame preamble, and in
θ(k) = θ(0), θ(T ), . . . , θ (N − 1)T (15) both the payload and CRC fields when the symbol s = 0
k=0
(corresponding to a pure upchirp) is to be sent. Moreover,
obtained taking N uniformly spaced samples of (14) in a sym- it may be used to generate also the downchirps required at the
bol interval. Assuming T = f1s = 2B1
, as previously stated, end of the preamble, by simply replacing5 sin(·) with − sin(·).
it results Most importantly, however, (17) and (18) are the basis to
   N −1
 N −1 π k
generate the modulated chirps for all symbols and all SFs,
θ(SF) (k) = k −1 + SF+1 . (16) as shown hereafter.
k=0 2 2
k=0
C. Step 3: Sampled Phase Term for Each SF and Each
The proof of (16) is reported in Appendix B. Modulation Symbol
Remark 1: One observes that (14) depends on the parame-
ters B and Ts . Given the relationship BTs = 2SF that holds for The paramount importance of (17) lays in that it is the
 2SF+1 −1
LoRa signals, it clearly appears that θ(t) depends, ultimately, basis to generate the sampled phase θ(SF) (s, k) for
k=0
on two of the three parameters B, Ts and SF. In (16), all modulation symbols s ∈ S and all SFs. This is possible
the dependence on SF appears explicitly in the right-hand term, thanks to the following equation:
whereas the dependence on B is implicitly taken into account    
1   1  
by the time interval T = 2B between two consecutive samples. θ(SF) s, k = θ(12) D (k + 2s) mod 2SF+1
For the sake of conciseness, in the left-hand term of (16) we D
explicitly highlighted in the superscript only the dependence 1  
on SF, which is the key parameter for the following discussion. − θ(12) 2sD , (19)
D
Having chosen fs = 2B, and recalling that BTs = M ,
it results N = fs Ts = 2M , that is, N = 2SF+1 . As it is with k = 0, 1, . . . , 2SF+1 − 1.
evident, N increases for increasing values of SF, reaching its 5 Given (14), the downchirp is obtained by simply taking −θ(t). Clearly,
maximum Nmax = 8192 for SF = 12. this impacts only on the sign of sin(·), which is an odd function.

Authorized licensed use limited to: INP TOULOUSE. Downloaded on June 22,2022 at 12:56:27 UTC from IEEE Xplore. Restrictions apply.
362 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 21, NO. 1, JANUARY 2022

V. D EMODULATION OF L O R A S IGNALS :
F REQUENCY D OMAIN I NTERPRETATION
In order to investigate the basics of the demodulation
method adopted by LoRa devices, let us assume that the
received signal is not affected by distortion, noise and inter-
ference. This means that, with a view to providing an insight
into the demodulation technique, the presence of receive filters
is not considered in this section. The receiver architecture and
the impact of filtering will be discussed, instead, in Sec.VI.
List. 1. MATLAB function for the generation of modulated phases.
We assume, moreover, that perfect carrier-, symbol-, and
frame-synchronizations have been established, so that the
receiver, after the RF to baseband conversion, exactly recovers
the complex envelope (8) in each symbol interval.
As observed in [12]–[14], in the n-th symbol interval
the receiver computes the twisted (dechirped) signal by
multiplying the received complex envelope (8) with the
complex-conjugate of a pure upchirp,6 whose phase is given
in (14). In particular, in the n-th symbol interval the receiver
computes:
rn (t) = in (t)e−jθ(t−nTs ) (20)
which results in
rn (t)
 B
V0 ej2π M s(t−nTs ) , nTs ≤ t < nTs + Th ,
= B
V0 ej2π(−B+ M s)(t−nTs ) , nTs + Th ≤ t ≤ (n + 1)Ts .
(21)
Fig. 4. Transmitter: Modulated phase with s = 91 and SF=8.
Since we are considering the generic symbol interval
[nTs , (n + 1)Ts], the term Th in (21) is the offset with respect
to the beginning of the same interval. The proof of (21) is
The proof of (19) is reported in Appendix D. reported in Appendix F.
In any symbol interval n, for any SF and modulation It clearly appears that rn (t) is a phasor with frequency
symbol s, (19) provides the corresponding phase samples to be (1) B
• f0 = M s in the interval [nTs , nTs + Th ),
fed to the cos(·) and sin(·) terms in (12) and (13). This results (2) B
• f0 = (−B + M s) in the interval [nTs + Th ,
in a simple implementation of LoRa transmitters, which can
thus be realized according to a digital architecture exploiting (n + 1)Ts ].
the capabilities of modern DSP microprocessors. Taking the Continuous Time Fourier Transform Rn (f )
The MATLAB function reported in Listing 1 shows the of (21) and computing its modulus leads to (see Appendix G):
 
phase generation algorithm: the function receives the mod-  
Rn (f )
ulation symbol s and the spreading factor SF as inputs, and   
 (1) (1) Th (1)
provides the discrete-time modulated phase (19) as output. The = V0 Th sinc (f −f0 )Th e−j2π(f −f0 )(nTs + 2 ) e−j2πf0 nTs
core of the function is represented by lines 5, 6, 7 and 8, which  
(2)
implement (19) starting from the reference upchirp phase (17) +Th sinc (f −f0 )Th
derived in lines 3 and 4. 
(2) T (2) 
×e−j2π(f −f0 )(nTs +Th + 2 ) e−j2πf0 nTs 
h
As an example case, the above MATLAB function has (22)
been used to derive the sampled phase depicted in Fig.4 with with Th denoting the residual chirp duration after the folding.
a dotted line-style, which refer to s = 91 and SF = 8 It turns out that the amplitude spectrum (22) contains two
(the lower curve, with a dashed line-style, is commented (1) (2)
sinc(·) functions centred in f0 and f0 , which appear as
in Appendix D). This is the modulated phase that origi- well-separated spectral lines, whose positions in the frequency
nates the instantaneous frequency-offset shown in Fig.1 when
axis depends on the modulation symbol s. Figure 5 shows an
B = 500 kHz. example of spectrum computed in the case s = 91, SF = 8,
Remark 2: It is worth observing that the sequence of
B = 500 kHz (the same setting used to derive Figs.1 and 4).
Nmax = 8192 samples provided by (17) is symmetric, meaning
Remark 3: With reference to (22), it is worth observing that
that θ(12) (k) = θ(12) (213 − k), with k = 0, 1, . . . , 212 (see (1) (2)
• f0 ∈ [0, M−1
M B], whereas f0 ∈ [−B, −B + M−1 M B],
Appendix E for the proof). This nice property of (17), clearly
meaning that the peak of the first sinc(·) is always in
evident in Fig.2(b), allows reducing by half the size of the
LUT needed for its storage. 6 The complex-conjugate of a pure upchirp corresponds to a pure downchirp.

Authorized licensed use limited to: INP TOULOUSE. Downloaded on June 22,2022 at 12:56:27 UTC from IEEE Xplore. Restrictions apply.
PASOLINI: ON Lo, vol. 21, no. 1, january 2022Ra CSS MODULATION: SIGNAL PROPERTIES AND THEIR IMPACT 363

Fig. 6. Periodic repetition of the original spectrum with aliasing.

In this regard, the key point is the sampling frequency

adopted in Step 1 for the discrete-time representation of
(d)
the dechirped signal, which is suggested to be fs = B
[11], [13], [14], [35]. It is worth noticing that, surprisingly,
this choice does not meet the sampling theorem requirement.
Fig. 5. Amplitude spectrum. s = 91, SF = 8, B = 500 kHz.
In fact, as observed in our Remark 3, the two spectral
lines of (21) fall in the frequency interval [−B, M−1 M B],
which implies that the sampling frequency should be such that
(d)
the right half-plane of the frequency domain (blue plot fs ≥ B[1 + M−1 M ]  2B. The straightforward consequence
in Fig.5), whereas the peak of the second sinc(·) is always of the undersampling is that the sampled dechirped signal
in the left half-plane (red plot in Fig.5); obtained in Step 1 is affected by aliasing in the frequency-
• the distance between the two spectral lines is B, irrespec- domain.
tively of the symbol s; More precisely, the discrete-time signal obtained sampling
• the spectrum (22) of the dechirped signal in a symbol (d)
the dechirped signal (21) with fs = B has a periodic spec-
interval is much more concentrated than the spectrum trum
of the modulated signal, which spans (and exceeds) the ∞

whole [− B2 , B2 ] interval [14]. In fact, the dechirping Rn(p) (f ) = fs(d) Rn (f − ifs(d) ) (23)
results in a despreading operation, turning the distributed i=−∞
spectrum of (8) into two compact spectral lines; (d)
whose period fs = B equals the frequency interval that sepa-
• although more concentrated, the spectrum of the
rates the two spectral lines (see Fig.5), which thus overlap each
dechirped signal covers a wider frequency range than the
other. This phenomenon is depicted in Fig.6, which shows that
spectrum of the modulated signal (8). In fact, depending
the periodic repetition with period B of the original spectrum
on the symbol s, the two spectral lines might appear
B (represented by color-filled contours) makes the first rightward
wherever (although with a discrete step M and separated
M−1 repetition of the red spectral line exactly overlap with the blue
by B) in the interval [−B, M B], whose size is almost
one, and vice-versa. Clearly, the resulting spectrum (23) after
twice that of the modulated signal spectrum.
Step 1 is the “coherent” sum of the overlapping components,
Clearly, starting from the dechirped signal spectrum (22), (d)
scaled by fs .
it is straightforward to derive the transmitted symbol s, which The FFT operation, carried out in Step 2 of the demodu-
is directly related to the frequency bins in which the two lation procedure, provides the receiver with a sampled rep-
spectrum peaks appear. Thus, it is not surprising that the resentation of (23) in the frequency interval [0, B], which is
demodulation method described in the patent documents [11], highlighted in Fig.6 by means of the coloured background box.
[12] is based on a numerical algorithm for the detection It is evident, now, that the spectral line that appears in such
of spectrum peaks. In particular, the demodulator described interval arises as the superposition (i.e., aliasing) of the two
in [11], [12] works as follows: spectral lines of the continuous-time dechirped signal (21).
• Step 1. It obtains the sampled version of the In Step 3 of the demodulation procedure, given the relation
continuous-time dechirped signal (21) with a properly (1) B
(d)
f0 = M s, it is straightforward for the demodulator to derive
chosen sampling frequency fs ; the transmitted symbol s by locating the frequency bin in
• Step 2. It performs the FFT of the sampled dechirped which the peak of this spectral line falls.
signal in each symbol time; (p) (1)
As far as the value A = |Rn (f0 )| of the useful peak is
• Step 3. It derives the modulation symbol by detecting the concerned, in Appendix H we show that it is an increasing
bin in which the peak of the FFT modulus appears. function of SF through the following simple equation:
As pointed out earlier, however, the spectrum of the
A = V0 2SF . (24)
continuous-time dechirped signal has two, non-symmetric,
peaks, in both the negative and positive semi-axes. So, where Equation (24) shows under a new perspective why higher
does the single peak mentioned by [11], [12] come from? SFs provide better performance in terms of communication

Authorized licensed use limited to: INP TOULOUSE. Downloaded on June 22,2022 at 12:56:27 UTC from IEEE Xplore. Restrictions apply.
364 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 21, NO. 1, JANUARY 2022

Fig. 8. Direct-conversion receiver. PA = power amplifier, LP = low-pass,

ADC = analog-to-digital converter.

thus deriving the in-phase and quadrature components p(r) (t)

and q (r) (t) of the received signal in a single down-conversion
step. The superheterodyne architecture, where RF signals are
down-converted to the baseband after a preliminary conversion
Fig. 7. Amplitude spectrum of the sampled dechirped signal. SF=8, to IF, is the well-known and widely adopted alternative,
Es
B = 500kHz, s = 91, V0 = 1, A = 256, N = 8 dB. although direct conversion has obvious advantages in terms
0

of power consumption, cost and hardware dimension, which

are fundamental aspects for the LoRa technology.
reliability. What’s more, (24) shows that A scales exponen- Irrespectively of the architecture adopted to derive p(r) (t)
tially with SF. and q (r) (t), either homodyne or superheterodyne, such base-
Remark 4: The overlapping of the two spectral components band signals are then sampled by ADCs, whose outputs are
due to the undersampling of (21) is extremely beneficial for then provided to a digital signal processing unit, such as
the demodulator, as the amplitude of the resulting spectral a DSP, an FPGA or an ASIC, where numerical techniques
line is higher than those of the two original ones. Clearly, this are implemented to complete the demodulation process.
increases the probability that its peak is detected even in the Taking as a reference the direct-conversion receiver shown
presence of noise and interference. in Fig.8, in this section we focus our attention on the low pass
This aspect can be clearly appreciated observing Fig.7, (LP) digital filters implemented in the numerical stage, which
which shows the FFT modulus of the discrete-time dechirped are highlighted as coloured blocks in the figure. The same
signal (that is, the input of Step 3), both in the presence digital filters are however implemented also in the baseband
and in the absence of AWGN noise, when s = 91, SF = 8 stage of superheterodyne receivers, which makes the following
and B = 500 kHz. The noise-impaired spectrum refers, discussion absolutely general.
Es
in particular, to the case N 0
= 8 dB, with Es denoting Before going into the details of our analysis, let us recall
the energy of the received chirp and N0 representing the the different objectives of the analog and digital filters shown
single-sided power spectrum of the noise. in Fig.8, which is instrumental to the clarity of our discussion.
One observes the existence of significant unwanted spec- The analog filters are in charge of removing out-of-band RF
tral peaks originated by the additive noise (which under- interferers (→ the RF analog filter) as well as selecting the
went the dechirping as well), that might mislead the useful channel (→ the LP analog filters), also removing the
receiver. By chance, in the specific case depicted in Fig.7 the noise and avoiding the occurrence of spectral aliasing with
noise also increased the useful spectrum peak, thus mitigating adjacent channels after the analog-to-digital conversion.
its detrimental impact. Considering the fact that the baseband stage of a LoRa
Figure 7 also shows that the peak of the uncorrupted spectral receiver is certainly equipped with a digital signal processing
line is A = 256, which is the expected value, following (24), unit, which is in charge of the FFT operation required for
as it has been obtained implementing the demodulation pro- the demodulation, LP digital filters can be easily implemented
cedure with MATLAB in the case V0 = 1 and SF = 8. as well (see the coloured blocks in Fig.8) with no addi-
tional hardware complexity, having the specific purpose of
VI. BASEBAND D IGITAL F ILTER D ESIGN reducing as much as possible the noise contribution. This is
Modern receivers are implemented combining analog and extremely beneficial, as it allows to design the analog RF
digital signal processing: the former takes place in the RF stage and LP filters with the only purpose of removing the adjacent
and, if any, in the intermediate frequency (IF) stage, whereas channel interferers, which are at least one guard-band apart.
numerical techniques are adopted in the baseband stage, It follows that, thanks to the presence of the LP digital filters,
where signals are represented through their sampled versions less frequency-selective (therefore, less complicated from a
obtained by means of analog-to-digital converters (ADCs). circuital point of view) analog RF and LP filters can be
In this regard, Fig.8 shows the architecture of a implemented.
direct-conversion (also called zero-IF or homodyne) receiver, Clearly, the counterpart is that the LP digital filters,
which converts the RF analog signal directly to the baseband, in charge of removing the out-of-band noise, must be very

Authorized licensed use limited to: INP TOULOUSE. Downloaded on June 22,2022 at 12:56:27 UTC from IEEE Xplore. Restrictions apply.
PASOLINI: ON Lo, vol. 21, no. 1, january 2022Ra CSS MODULATION: SIGNAL PROPERTIES AND THEIR IMPACT 365

Fig. 9. Filter design mask.

selective, resembling as much as possible to ideal filters,

that is, non-distorting filters with a pass-band equal to the Fig. 10. Symbol error rate. B = 125 kHz, fpass = 62.5 kHz.
bandwidth B299  B2 of the baseband signals p(r) (t) and
q (r) (t). As it turns out, the reduced circuital complexity of
analog filters is traded with the increased complexity of the • Pw = N0 B denotes the noise power within the nominal
software implementation of digital filters. signal bandwidth B.
As it is known, non-distorting digital filters can be imple- Such reference performance curves, obtained assuming ideal
mented using finite impulse response (FIR) design techniques, receive filters with a pass-band equal to the nominal signal
with a selectivity that improves as the number of filter bandwidth, are plotted in Fig.10 with a dashed line-style.
coefficients increases. Improving the selectivity, which has a Considering LoRa signals with B = 125kHz, the same curves
beneficial impact on the receiver performance, has, however, were obtained with our simulator by adopting the following
two major drawbacks: settings for the two LP digital filters:
• an increased computational power (that is, a more pow- B
• fpass = 2 = 62.5 kHz, which is coincident with the
erful hardware) may be required, as the number of nominal bandwidth of the baseband signals p(r) (t) and
multiplications and sums carried out in the unit time q (r) (t) at the filters’ input ports,
by the two digital filters increases with the number of • Apass = 0.01 dB, which means that the amplitude distor-
coefficients, tion introduced by the filters’ ripple within the pass-band
• an increased energy consumption, due to the augmented is practically negligible,
computational effort. • fstop = 64 kHz, which corresponds to very selective
Both issues are quite critical for LoRa end nodes, which filters, as the width of the transition band is only 1.5 kHz.
are required to cost only a few dollars and last for many This is consistent with the attempt to approximate the
years. Therefore, the receiver design should be carried out ideal filters considered in [15], [16],
considering the relation between the receiver performance and • Astop = 30dB, which appears reasonable for practical
the number of filter coefficients, which is the focus of this filters. Clearly, given the hypothesis of ideal filters,
section. [15], [16] assume Astop = +∞,
We addressed this topic by implementing a MATLAB (ADC) (ADC)
• fs = 250 ksamples
s , with fs denoting the sam-
simulator of a complete LoRa transceiver (transmitter + pling frequency adopted by the ADCs, which represents
receiver) working in an AWGN channel. At the receiver side, the filters’ clock.
in particular, we implemented the two linear-phase LP FIR As shown in Fig.10, when such filters (each requiring
filters depicted in Fig.8, whose characteristics, namely, 409 coefficients) are implemented in the LoRa simulator,
• the pass-band width fpass , the resulting symbol error rates are almost coincident with
• the pass-band ripple Apass , the reference ones for each SF. This is immediately evident
• the lower stop-band edge fstop , in Fig.10 for SF = 8, 9, 10 and 11, for which only the
• the stop-band attenuation Astop , reference and the simulated curves are reported. As far as
have been properly tuned, as discussed in the following. the cases SF = 7 and SF = 12 are concerned, the practical
The reader is referred to Fig.9 for an explicative, pictorial coincidence between the reference curves and the simulated
representation of the above mentioned parameters. ones is also true, although less evident in Fig.10, which shows
As a first step, we validated the simulator by retrieving the bunches of curves related to the experienced symbol error
results reported in [15] and [16], which refer to the symbol rates also for less selective filters, with fstop = 64kHz, 68kHz,
error rate (i.e., the chirp error rate) as a function of the signal- 72kHz, . . . , 98kHz.
to-noise ratio SNR = PPws , where In particular, still assuming fpass = 62.5 kHz,
• Ps denotes the average signal power, Apass = 0.01 dB and Astop = 30 dB, the two bunches

Authorized licensed use limited to: INP TOULOUSE. Downloaded on June 22,2022 at 12:56:27 UTC from IEEE Xplore. Restrictions apply.
366 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 21, NO. 1, JANUARY 2022

TABLE II algorithm for the numerical computation of the signal samples

T RADE -O FF B ETWEEN F ILTER S ELECTIVITY at the transmitter side. Then, we showed that the effectiveness
AND R ECEIVER P ERFORMANCE
of the demodulation technique depends on the fact that the
sampling frequency at the receiver is forced to violate the
sampling theorem requirement. Finally, we showed that the
number of coefficients of the digital filters implemented in the
baseband stage of the receiver might be significantly reduced,
at the cost of an almost negligible increase of the signal-
to-noise ratio required for a given performance level.

A PPENDIX A
P ROOF OF (6)
Proof: Given (4), it follows that
 Ts
θ(s, Ts ) = 2π Δf (s, ξ)dξ = 0.
0
Being, of course,
 0
of curves for SF = 7 and SF = 12 highlight the performance θ(s, 0) = 2π Δf (s, ξ)dξ = 0
degradation that is to be expected passing from fstop = 64 kHz 0
(leftmost simulated curve of each bunch) to fstop = 98 kHz the proof of (6) is immediately obtained.
(rightmost simulated curve of each bunch), which corresponds,
for both SF = 7 and SF = 12, to an SNR loss of about A PPENDIX B
1dB for a target symbol error rate of 10−3 . On the other P ROOF OF (16)
hand, as shown in Table II, the benefit in terms of number Proof: Given (14) and (15), it is:
of coefficients required to implement each of the two filters  N −1    N −1
k2 T 2
is quite relevant, as it passes from 409 for fstop = 64 kHz to θ(k) = πB −kT + .
k=0 Ts k=0
only 17 for fstop = 98 kHz.
1
Clearly, intermediate choices are possible, which are pre- Having assumed fs = 2B, one obtains T = 2B and therefore:
sented in Table II. For instance, passing from fstop = 64 kHz  N −1    N −1
k k2
to fstop = 68 kHz allows saving nearly 300 coefficients for θ(k) = πB − +
k=0 2B 4B 2 Ts k=0
each filter with an almost insignificant SNR loss for the target    N −1
symbol error rate 10−3 . π k
(ADC) = k −1 + .
Comments on fs and f (d) : In this section, we assumed 2 2BTs k=0
(ADC)
that the sampling frequency fs adopted by the ADCs Since BTs = M = 2SF , and explicitly showing the depen-
fully meets the Shannon-Hartley sampling theorem. This is dence on SF, (16) is immediately derived.
a mandatory condition to design digital filters capable of pre-
serving the useful signal while reducing the out-of-band noise. A PPENDIX C
However, in Sec.V we pointed out that the sampling frequency P ROOF OF (18)
f (d) = B, which purposely does not meet the sampling Proof: Given the reference sampled phase (17), one
theorem requirement, must be adopted in the demodulation considers its decimated version, obtained taking 1 sample
process (Sec.V, Step 1). This means that the demodulator must every D = 212−SF samples. It results
include a downsampling stage, which reduces the sampling    2D −1
13
(ADC)   2D13 −1 
frequency from fs to f (d) = B immediately after the π Dk
filters. θ(12) (Dk) = Dk −1 + 13 .
k=0 2 2
In order to keep the downsampler as simple as possible, it is k=0
(ADC) 213
convenient to choose fs as an integer multiple of f (d) , Observing that D =2 SF+1
, one obtains
(ADC) (d)
that is, fs = Lf = LB, with L ∈ Z and L > 1.    2SF+1 −1
 2SF+1 −1 π k
In such a case, the downsampler needs only to select one
θ(12) (Dk) =D k −1 + SF+1 .
sample every L. In the numerical results, given the choice k=0 2 2
(ADC) k=0
fs = 250 ksamples
s for the considered LoRa signals with
Recalling (16) it follows
B = 125 kHz, we assumed L = 2.
 2SF+1 −1  2SF+1 −1
θ(12) (Dk) = D θ(SF) (k) ,
VII. C ONCLUSION k=0 k=0

In this paper, we analytically investigated the generation and, ultimately

 2SF+1 −1 1  (12) 2SF+1 −1
and the demodulation of LoRa signals, as well as the trade-off
θ(SF) (k) = θ (Dk) .
between the receiver performance and the number of coeffi- k=0 D k=0
cients of the receive digital filters. We firstly provided a simple

Authorized licensed use limited to: INP TOULOUSE. Downloaded on June 22,2022 at 12:56:27 UTC from IEEE Xplore. Restrictions apply.
PASOLINI: ON Lo, vol. 21, no. 1, january 2022Ra CSS MODULATION: SIGNAL PROPERTIES AND THEIR IMPACT 367

A PPENDIX D dotted line-style. One observes, by the way, that both phases
P ROOF OF (19) shown in Fig.4 provide the right frequency sweep, which is
Proof: For a given SF, (16) provides the (sampled) phase the one depicted in Fig.1. Nonetheless, the dashed curve is not
term of a pure upchirp, whose instantaneous frequency-offset compliant with the LoRa specifications.
sweeps the entire interval [− B2 , B2 ] starting from its lowest Plugging (18) in (26), it finally results:
   
boundary − B2 (see Fig.2(a)). A chirp modulated by a symbol   1  
s is different from a pure upchirp in that its instantaneous θ(SF) s, k = θ(12) D (k + 2s) mod 2SF+1
D
frequency-offset sweeps the same interval starting from a  
higher value, given by − B2 + MB
s, then it wraps around to − B2 1
− θ(12) 2sD ,
B
as soon as it reaches 2 and keeps on increasing until the end D
of the symbol interval (see Fig.1). with k = 0, 1, . . . , 2SF+1 − 1.
Considering the reference symbol interval [0, Ts ], the instan-
taneous frequency-offset of a chirp modulated by a symbol s A PPENDIX E
can thus be considered as the result of a time advance S YMMETRY OF THE R EFERENCE U PCHIRP P HASE
(that is, a left shift along the time axis), whose entity
depends on s, of the instantaneous frequency-offset of a It must be proved that θ(12) (213 − k) = θ(12) (k) with
pure upchirp. This modelling is complete on condition that k = 0, 1, . . . , 212 .
the symbol time interval is given the circularity property, Proof: Starting from (17) it is
 
so that the translated instantaneous frequency-offset wraps (12) 13 13 π 213 − k
around “horizontally” in t = 0, making it enter from the θ (2 − k) = (2 − k) −1 + .
2 213
right boundary of the symbol interval what exits from the left
boundary. After easy manipulations of the right-hand term, one obtains:
In particular, starting from the phase term (16) of a pure π  k 
θ(12) (213 − k) = k − 1 + 13
upchirp, a new phase term can be derived as follows, which 2 2
provides the frequency sweep associated to a certain modula- Recalling (17) it immediately follows:
tion symbol s:
 θ(12) (213 − k) = θ(12) (k).
 2SF+1 −1
(SF) SF+1
θ (k + 2s) mod 2 . (25) Clearly, both terms of the above equality are sound
k=0
only provided that their arguments are meaningful, which
Equation (25) incorporates both the time advance, given means 0 ≤ k ≤ 213 . Such interval includes the range
by 2s, and the wrap-around feature, given by the mod oper- k = 0, 1, . . . , 212 of interest for this proof.
ation. One observes, in particular, that since the number of
modulation symbols is M , whereas the number of samples in A PPENDIX F
a chirp interval is 2M (see, Sec.IV-A), the number of samples P ROOF OF (21)
corresponding to a time advance of s is 2s.
Although (25) provides the right frequency sweep for Proof: Given (8), in the n-th symbol interval
any given symbol s, it lacks of the fundamental prop- [nTs , (n + 1)Ts ] the receiver obtains:
erty of null phase at the boundaries of symbol intervals rn (t) = in (t)e−jθ(t−nTs )
(Property 1, Sec.III). This is clearly evident in Fig.4, where (n)
,t) −jθ(t−nTs )
the dashed curve represents the phase term given by (25) in = V0 ejθ(s e , (27)
the case s = 91 and SF = 8. where θ(s(n) , t) can be easily derived from (3) as in (28),
The null phase at the boundaries of the symbol interval shown at the bottom of the page.
can be simply obtained by compensating (i.e., removing) Plugging (28) and (14) in (27) immediately leads to (21).
the vertical offset θ(SF) (2s) exhibited by (25) for k = 0
(see Fig.4), thus obtaining:
     
(SF) (SF) SF+1 A PPENDIX G
θ s, k = θ (k + 2s) mod 2 P ROOF OF (22)
 
−θ(SF) 2s (26) Proof: Equation (21) can be rewritten as
(1)
(t−nTs )
with k = 0, 1, . . . , 2SF+1 − 1. In the example case s = 91 and rn (t) = V0 ej2πf0 gTh (t − nTs )
(2)
j2πf0 (t−nTs )
SF = 8, the resulting phase term is shown in Fig.4 with a +V0 e gTh (t − nTs − Th ). (29)


2π[− B2 (t − nTs ) + M
B
s(t − nTs ) + 2T
B
(t − nTs )2 ], nTs ≤ t < nTs + Th ,
θ(s(n) , t) = 3B
s
(28)
2π[− 2 (t − nTs ) + M s(t − nTs ) + 2Ts (t − nTs )2 ], nTs + Th ≤ t ≤ (n + 1)Ts
B B

Authorized licensed use limited to: INP TOULOUSE. Downloaded on June 22,2022 at 12:56:27 UTC from IEEE Xplore. Restrictions apply.
368 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 21, NO. 1, JANUARY 2022

   Th

(1)
Rn (f ) = V0 δ(f − f0 )e−j2πf nTs ∗ Th sinc(f Th )e−j2πf (nTs + 2 )
   T 
(2)
+ V0 δ(f − f0 )e−j2πf nTs ∗ Th sinc(f Th )e−j2πf (nTs +Th + 2 )
h

  (1) Th (1)
(1)
= V0 Th sinc (f − f0 )Th e−j2π(f −f0 )(nTs + 2 ) e−j2πf0 nTs
  (2) T (2)
(2)
+V0 Th sinc (f − f0 )Th e−j2π(f −f0 )(nTs +Th + 2 ) e−j2πf0 nTs .
h
(30)
 (1) Th (1)
(1)
Rn(p) (f ) = V0 B Th sinc[(f − f0 )Th ]e−j2π(f −f0 )(nTs + 2 ) e−j2πf0 nTs
(2) T (2)

(2)
+Th sinc[(f − f0 − B)Th ]e−j2π(f −f0 −B)(nTs +Th + 2 ) e−j2πf0
h nTs
. (33)
 (1) Th (1)
(1)
Rn(p) (f ) = V0 B Th sinc[(f − f0 )Th ]e−j2π(f −f0 )(nTs + 2 ) e−j2πf0 nTs
(1) T (2)

(1)
+Th sinc[(f − f0 )Th ]e−j2π(f −f0 )(nTs +Th + 2 ) e−j2πf0 nTs .
h
(34)

Taking the continuous Fourier transform and denoting with ACKNOWLEDGMENT

* the linear convolution, (30), shown at the top of the The author is in debt with Prof. Marco Chiani for his
page, is immediately obtained, with δ(·) denoting the Dirac’s inspiring paper [14] and for the fruitful and encouraging
delta. discussions on LoRa. The author is also deeply grateful to
A PPENDIX H Prof. Oreste Andrisano for helpful suggestions.
P ROOF OF (24)
R EFERENCES
Proof: In Appendix G we showed that the spectral
(1) [1] K. Ashton. (2009). That ’Internet of Things’ Thing.
component of (22) centred in f0 arises from the term
  (1) Th (1)
Accessed: Jul. 6, 2021. [Online]. Available: https://www.rfidjournal.com/
(1)
V0 Th sinc (f − f0 )Th e−j2π(f −f0 )(nTs + 2 ) e−j2πf0 nTs that-internet-of-things-thing
[2] U. Raza, P. Kulkarni, and M. Sooriyabandara, “Low power wide area
(31) networks: An overview,” IEEE Commun. Surveys Tuts., vol. 19, no. 2,
pp. 855–873, 2nd Quart., 2017.
(2)
whereas the spectral component centred in arises from
f0 [3] K. Mekki, E. Bajic, F. Chaxel, and F. Meyer, “Overview of cellular
LPWAN technologies for IoT deployment: Sigfox, LoRaWAN, and NB-
the term
  (2) T (2)
IoT,” in Proc. IEEE Int. Conf. Pervas. Comput. Commun. Workshops
(2)
V0 Th sinc (f −f0 )Th e−j2π(f −f0 )(nTs +Th + 2 ) e−j2πf0 nTs . (PerCom Workshops), Mar. 2018, pp. 197–202.
h

[4] (2020). LoRa Alliance. Accessed: Jul. 6, 2021. [Online]. Available:

(32) https://lora-alliance.org/
[5] O. Khutsoane, B. Isong, and A. M. Abu-Mahfouz, “IoT devices and
Recalling (23) and focusing the attention on the frequency applications based on LoRa/LoRaWAN,” in Proc. 43rd Annu. Conf.
IEEE Ind. Electron. Soc. (IECON), Oct. 2017, pp. 6107–6112.
interval [0, B], we observe that, owing to the undersampling, [6] J. Haxhibeqiri, E. De Poorter, I. Moerman, and J. Hoebeke, “A survey of
(32) is periodically repeated with period B and overlaps LoRaWAN for IoT: From technology to application,” Sensors, vol. 18,
(1) no. 11, p. 3995, Nov. 2018.
with (31) around f0 , generating the spectral component
[7] G. Pasolini et al., “Smart city pilot project using LoRa,” in Proc. Eur.
given by (33), shown at the top of the page, for f ∈ [0, B]. Wireless 24th Eur. Wireless Conf., 2018, pp. 1–6.
(2) (1)
Observing that f0 + B = f0 , (34) shown at the top [8] G. Pasolini et al., “Smart city pilot projects using LoRa and
of the page is immediately derived. IEEE802.15.4 technologies,” Sensors, vol. 18, no. 4, p. 1118, Apr. 2018.
(1) [9] (2020). Semtech Corporation. Accessed: Jul. 6, 2021. [Online]. Avail-
Finally, computing the value of (34) for f = f0 gives able: https://www.semtech.com/lora
(1) [10] (2020). LoRaWAN Specifications. Accessed: Jul. 6, 2021. [Online].
Rn(p) (f0 )
 (1) (2)
 Available: https://lora-alliance.org/resource-hub/lorawanr-specification-
v103
= V0 B Th e−j2πf0 nTs + Th e−j2πf0 nTs [11] (2020). EP2763321A1. Accessed: Jul. 6, 2021. [Online]. Available:
 (1) (1)
 https://patents.google.com/patent/EP2763321A1
= V0 B Th e−j2πf0 nTs + Th e−j2π(f0 −B)nTs [12] (2020). EP2449690B1. Accessed: Jul. 6, 2021. [Online]. Available:
 (1) (1)
 https://patents.google.com/patent/EP2449690B1/en
= V0 B Th e−j2πf0 nTs + Th e−j2πf0 nTs ej2πBnTs [13] L. Vangelista, “Frequency shift chirp modulation: The LoRa modu-
 (1) (1)
 lation,” IEEE Signal Process. Lett., vol. 24, no. 12, pp. 1818–1821,
Dec. 2017.
= V0 B Th e−j2πf0 nTs + Th e−j2πf0 nTs ej2πMn [14] M. Chiani and A. Elzanaty, “On the LoRa modulation for IoT: Waveform
  (1) properties and spectral analysis,” IEEE Internet Things J., vol. 6, no. 5,
= V0 B Th + Th e−j2πf0 nTs pp. 8463–8470, May 2019.
(1)
[15] O. Afisiadis, M. Cotting, A. Burg, and A. Balatsoukas-Stimming,
= V0 BTs e−j2πf0 nTs “On the error rate of the LoRa modulation with interference,” IEEE
(1) Trans. Wireless Commun., vol. 19, no. 2, pp. 1292–1304, Feb. 2019.
−j2πf0 nTs
= V0 2SF e . (35) [16] G. Ferre and A. Giremus, “LoRa physical layer principle and perfor-
mance analysis,” in Proc. 25th IEEE Int. Conf. Electron., Circuits Syst.
Taking the modulus of (35) leads to (24). (ICECS), Dec. 2018, pp. 65–68.

Authorized licensed use limited to: INP TOULOUSE. Downloaded on June 22,2022 at 12:56:27 UTC from IEEE Xplore. Restrictions apply.
PASOLINI: ON Lo, vol. 21, no. 1, january 2022Ra CSS MODULATION: SIGNAL PROPERTIES AND THEIR IMPACT 369

[17] T. Elshabrawy and J. Robert, “Closed-form approximation of LoRa [30] D. Bankov, E. Khorov, and A. Lyakhov, “Mathematical model of
modulation BER performance,” IEEE Commun. Lett., vol. 22, no. 9, LoRaWAN channel access with capture effect,” in Proc. IEEE 28th
pp. 1778–1781, Sep. 2018. Annu. Int. Symp. Pers., Indoor, Mobile Radio Commun. (PIMRC),
[18] H. Mroue, A. Nasser, B. Parrein, S. Hamrioui, E. Mona-Cruz, and Oct. 2017, pp. 1–5.
G. Rouyer, “Analytical and simulation study for LoRa modulation,” in [31] C. Pham, A. Bounceur, L. Clavier, U. Noreen, and M. Ehsan, “Inves-
Proc. 25th Int. Conf. Telecommun. (ICT), Jun. 2018, pp. 655–659. tigating and experimenting interference mitigation by capture effect in
[19] A. Marquet, N. Montavont, and G. Z. Papadopoulos, “Investigating LoRa networks,” in Proc. 3rd Int. Conf. Future Netw. Distrib. Syst.,
theoretical performance and demodulation techniques for LoRa,” in New York, NY, USA, Jul. 2019.
Proc. IEEE 20th Int. Symp. World Wireless, Mobile Multimedia Netw. [32] (2020). Semtech Corporation. AN1200.22 LoRa Modulation Basics,
(WoWMoM), Jun. 2019, pp. 1–6. May 2015. Accessed: Jul. 6, 2021. [Online]. Available: http://wiki.
[20] T. T. Nguyen, H. H. Nguyen, R. Barton, and P. Grossetete, “Efficient lahoud.fr/lib/exe/fetch.php?media=an1200.22.pdf
design of chirp spread spectrum modulation for low-power wide-area [33] A. Augustin, J. Yi, T. Clausen, and W. Townsley, “A study of LoRa:
networks,” IEEE Internet Things J., vol. 6, no. 6, pp. 9503–9515, Long range & low power networks for the Internet of Things,” Sensors,
Dec. 2019. vol. 16, no. 9, p. 1466, Sep. 2016.
[21] T. Elshabrawy and J. Robert, “Interleaved chirp spreading LoRa-based [34] N. Sornin and A. Yegin. (2017). LoRaWAN Specification.
modulation,” IEEE Internet Things J., vol. 6, no. 2, pp. 3855–3863, Accessed: Jul. 6, 2021. [Online]. Available: https://lora-alliance.org/wp-
Apr. 2019. content/uploads/2020/11/lorawantm_specification_-v1.1.pdf
[22] T. Elshabrawy, P. Edward, M. Ashour, and J. Robert, “On the different [35] R. Ghanaatian, O. Afisiadis, M. Cotting, and A. Burg, “LoRa digital
mathematical realizations for the digital synthesis of LoRa-based modu- receiver analysis and implementation,” in Proc. IEEE Int. Conf. Acoust.,
lation,” in Proc. Eur. Wireless, 25th Eur. Wireless Conf., 2019, pp. 1–6. Speech Signal Process. (ICASSP), May 2019, pp. 1498–1502.
[23] A. Thomas and N. V. Eldhose, “Performance evaluation of chirp spread
spectrum as used in LoRa physical layer,” in Proc. IEEE Int. Conf. Syst.,
Comput., Autom. Netw. (ICSCAN), Mar. 2019, pp. 1–6.
[24] B. Dunlop, H. H. Nguyen, R. Barton, and J. Henry, “Interference analysis
for LoRa chirp spread spectrum signals,” in Proc. IEEE Can. Conf.
Electr. Comput. Eng. (CCECE), May 2019, pp. 1–5. Gianni Pasolini (Member, IEEE) received the M.Sc.
[25] J. M. de Souza Sant’Ana, A. Hoeller, R. D. Souza, H. Alves, degree in telecommunications engineering and the
and S. Montejo-Sánchez, “LoRa performance analysis with superposed Ph.D. degree in electronic engineering and com-
signal decoding,” IEEE Wireless Commun. Lett., vol. 9, no. 11, puter science from the University of Bologna, Italy,
pp. 1865–1868, Nov. 2020. in 1999 and 2003, respectively. He is currently an
[26] U. Noreen, L. Clavier, and A. Bounceur, “LoRa-like CSS-based PHY Assistant Professor with the Department of Electri-
layer, capture effect and serial interference cancellation,” in Proc. 24th cal, Electronic and Information Engineering (DEI),
Eur. Wireless Conf., 2018, pp. 1–6. University of Bologna, Italy, where he has been
[27] P. Edward, A. Muhammad, S. Elzeiny, M. Ashour, T. Elshabrawy, and holding the professorship of several courses in the
J. Robert, “Enhancing the capture capabilities of LoRa receivers,” in telecommunications area since 2003 and holds the
Proc. Int. Conf. Smart Appl., Commun. Netw. (SmartNets), Dec. 2019, “Signal Processing” course with the Engineering
pp. 1–6. School. He has been involved in several European initiatives on wireless com-
[28] C. Pham and M. Ehsan, “Dense deployment of LoRa networks: Expec- munications, such as COST actions and Networks of Excellence (Newcom and
tations and limits of channel activity detection and capture effect for Newcom#). His research interests include wireless communication systems,
radio channel access,” Sensors, vol. 21, no. 3, p. 825, Jan. 2021. the Internet of Things, digital signal processing, and THz communications.
[29] R. B. Sorensen, N. Razmi, J. J. Nielsen, and P. Popovski, “Analysis of He is the Vice Chair of the Joint ComSoc/VTS Italian Chapter and an
LoRaWAN uplink with multiple demodulating paths and capture effect,” Associate Editor of the IEEE O PEN J OURNAL OF THE C OMMUNICATIONS
in Proc. IEEE Int. Conf. Commun. (ICC), May 2019, pp. 1–6. S OCIETY (IEEE OJ-COM).

Authorized licensed use limited to: INP TOULOUSE. Downloaded on June 22,2022 at 12:56:27 UTC from IEEE Xplore. Restrictions apply.