1818 IEEE SIGNAL PROCESSING LETTERS, VOL. 24, NO.

12, DECEMBER 2017

Frequency Shift Chirp Modulation:

The LoRa Modulation
Lorenzo Vangelista , Member, IEEE

Abstract—Low power wide area networks (LPWAN) are emerg- factor S, log2 (S) bits define f0 ,” i.e., the initial frequency shift,
ing as a new paradigm, especially in the field of Internet of Things but there is no explanation of how this is done.
(IoT) connectivity. LoRa is one of the LPWAN and it is gaining quite
Actually the LoRa modulation is often referred to as a “chirp
a lot of commercial traction. The modulation underlying LoRa is
patented and has never been described theoretically. The aim of modulation” [8]–[10]. A close inspection of LoRa reveals that
this letter is to give the first rigorous mathematical signal process- the information bearing element is the frequency shift at the
ing description of the modulation and demodulation processes. We beginning of the symbol and the chirp is similar to a kind of a
provide as well a theoretical derivation of the optimum receiver carrier. For this reason, in our opinion, LoRa is better described
entailing a low-complexity demodulation process, resorting to the
as a Frequency Shift Chirp Modulation (FSCM).
Fast Fourier Transform. We compare then the performance of the
LoRa modulation and the frequency-shift keying modulation both The rest of the paper is organised as follows. In Section II,
in an additive white gaussian noise channel and a frequency selec- we provide a description of the modulation process and identify
tive channel, showing the superiority of the LoRa modulation in the orthogonal signals basis characterising the modulation; in
the frequency selective channel. The results of this letter will en- Section III, we provide a description of the optimum demod-
able a further assessment of the LoRa based networks, much more
ulator and an efficient implementation of it making use of the
rigorous than what has been done until now.
Fast Fourier Transform; in Section IV, we provide the results
Index Terms—Chirp modulation, frequency shift chirp modula- of some computer simulations experiments on the link level
tion (FSCM), Internet of Things (IoT), LoRa, low power wide area performances, comparing as well the FSCM modulation with
networks (LPWAN).
an frequency-shift keying (FSK) modulation with the same car-
dinality. Finally in Section V, we draw the conclusions of the
I. INTRODUCTION paper.
OW power wide area networks (LPWAN) are emerging
L as a new paradigm, especially in the field of Internet of
Things connectivity [1]–[3].
II. FREQUENCY SHIFT CHIRP MODULATION
Let us suppose that the bandwidth of the channel we use for
LoRa is one of the LPWAN (see [3]) and it is gaining quite the transmission is B so we transmit a sample every T = B1 .
a lot of commercial traction. Strictly speaking, LoRa is the A symbol s(nTs ) is sent at the input of the modulator every
physical layer of the LoRaWAN system, whose specification Ts = 2SF · T . The symbol s(nTs ) is a real number formed using
is maintained by the LoRa Alliance.1 The LoRa modulation is a vector w(nTs ) of SF binary digits, with SF an integer param-
patented and has never been described theoretically. The patent eter called, in the context of LoRa, Spreading Factor (usually
[4], indeed, does not provide the details, in term of equations and taking values in {7, 8, 9, 10, 11, 12}) i.e.,
signal processing. Paper [5] gives a high level description of the
LoRa modulation, providing some basic equations and relying 
SF−1

on the intuition of the reader for the decoding process. Papers s(nTs ) = w(nTs )h · 2h . (1)
h=0
[6] and [7] get more in detail with the signal, modulation and
demodulation description, but still lack a mathematical, signal We can see that s(nTs ) takes values in {0, 1, 2, . . . , 2SF − 1}.
theory based definition of the modulation and demodulation The transmitted waveform, of duration Ts , for a certain
processes, partly because the analysis is limited to the analog s(nTs ) is then
domain. For example, in [7] it is said that “For a spreading
1
c(nTs + kT ) = √ ej2π [(s(n T s )+k ) m o d 2 SF ]k T 2BSF (2)
2SF
Manuscript received August 17, 2017; revised October 3, 2017; accepted
1
ej2π [(s(n T s )+k ) m o d ]
k
October 10, 2017. Date of publication October 13, 2017; date of current ver- =√ 2 SF 2 SF (3)
sion October 25, 2017. The associate editor coordinating the review of this 2 SF
manuscript and approving it for publication was Dr. Andre L. F. de Almeida.
The author is with the Department of Information Engineerng, University of for k = 0 . . . 2SF − 1.
Padova, Padova 35131, Italy (e-mail: [email protected]).
Color versions of one or more of the figures in this letter are available online We can see that the modulated signal is a chirp waveform, as
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/LSP.2017.2762960 the frequency increases linearly with k, which is the time index;
1 https://www.lora-alliance.org/ we note that each waveform differs from a base waveform having

1070-9908 © 2017 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.

Authorized licensed use limited to: ULAKBIM UASL - ANKARA UNIVERSITESI. Downloaded on January 09,2024 at 17:27:45 UTC from IEEE Xplore. Restrictions apply.
 VANGELISTA: FREQUENCY SHIFT CHIRP MODULATION: THE LORA MODULATION 1819

initial frequency equal to 0 by a frequency shift s(nTs ). That is 2

SF
−1
1 k
why we call it FCSM. + SF ej2π [i−q ] 2 SF
2
We remark that all of the analysis of the FCSM modula- k =2 SF −1−q
tion in this letter will remain in the discrete domain Z(T ) = 2 −1
1  j2π [i−q ] kSF
SF

{. . . , −3T, −2T, −T, 0, T, 2T, 3T, . . .}, i.e., the fundamental = SF e 2 . (8)
interval for the frequency analysis is [0, B = T1 ]. As a matter of 2
k =0
fact any signal in the discrete domain Z(T ) has a frequency rep-
Now we consider first the case of i − q a odd number. In this
resentation periodic with period B = T1 [12]. Therefore, if one
case, we can write the following:
prefers to have the FCSM described in the frequency interval  
[−B/2, B/2], e.g., to deal with the analytic signal, the signal c(nTs + kT )|s(n T s )=i , c(nTs + kT )|s(n T s )=q
base (4) just need to be multiplied by e−j2π 2 k T = −1k , with
B

−1
1  j2π [i−q ] kSF
SF−1
no consequences on the derivations and findings of the current 2
k + 2 SF−1

letter. = SF
e 2 + ej2π [i−q ] 2 SF
2
k =0

−1
1  j2π [i−q ] kSF
SF−1
2
A. On the Orthogonality of the FCSM Waveforms
+ ej2π [i−q ] ( 2 SF + 2 )
k 1
= SF
e 2
Having established the possible waveforms of the FCSM, we 2
k =0
investigate their orthogonality. To this end we need to check
−1
1  j2π [i−q ] kSF
SF−1
2
  k
= e 2 − ej2π [i−q ] 2 SF = 0.
c(nTs + kT )|s(n T s )=i , c(nTs + kT )|s(n T s )=q = 0 2SF
k =0

i = q, i, q ∈ {0, . . . , 2 − 1} SF
(4) If i − q is even then it can always be written as i − q = 2d · r,
  where d is an integer, such that 0 ≤ d < SF and r an odd integer.
c(nTs + kT )|s(n T s )=i , c(nTs + kT )|s(n T s )=q (5) We tackle first the subcase r = 1
 
2
SF
−1 c(nTs + kT )|s(n T s )=i , c(nTs + kT )|s(n T s )=q
= c(nTs + kT )|s(n T s )=i · c∗ (nTs + kT )|s(n T s )=q
2 −1 2 −1
1  j2π [i−q ] kSF 1  j2π [2 q ·r ] kSF
SF SF
k =0

2 −1
= SF e 2 = SF e 2

1  j2π [(i+k ) m o d
SF
2 2
] e−j2π [(q +k ) m o d ]
k k k =0 k =0
= e 2 SF 2 SF 2 SF 2 SF
2SF 2 −1
1  j2π r SF−q
SF
k =0 k

2 −1
= e 2

1  j2π [(i+k ) m o d
SF SF
2
2 SF −(q +k ) m o d 2 SF 2 SF]
k k =0
= e . (6)
2SF 2 −1 2 −1
1   j2π r h ·2SF−q
q SF−q
k =0 SF + p
= SF e 2
Equation (6) can be further elaborated, supposing (without 2 p=0
h=0
loss of generality) i > q, by splitting it in three parts, getting rid
2 −1 2 −1
1   j2π r SF−q
q SF−q
of the mod operator in in each of them. p
= SF e 2
  2 p=0
h=0
c(nTs + kT )|s(n T s )=i , c(nTs + kT )|s(n T s )=q
2 −1 2 −1
1   j2π r SF−q
q SF−q
p
−1−i
2 SF = SF e =0
1 k
2

= ej2π [(i+k )−(q +k )] 2 SF 2 p=0

h=0
2SF
k =0 2 SF−q −1 p

−1−q
2 SF
since p=0 ej2π r 2 SF−q = 0 being r odd.
1 We tackle now the last remaining case r = 1.
ej2π [(i+k −2 )−(q +k ) ] 2 SF
SF k
+  
2SF
k =2 SF −i c(nTs + kT )|s(n T s )=i , c(nTs + kT )|s(n T s )=q
2
SF
−1
1 2 −1 2 −1
1  j2π [i−q ] kSF 1  j2π 2 q ·r kSF
SF SF
ej2π [(i+k −2 )−(q +k −2 )] 2 SF
SF SF k
+ (7)
2SF = SF e 2 = SF e 2
k =2 SF −q 2 2
k =0 k =0
−1−i
2 SF 2
SF
−1 2
q
−1 2 SF−q
−1
1 k 1 k 1 h ·2 SF−q + p
= ej2π [i−q ] 2 SF = ej2π 2 SF−q = ej2π 2 SF−q
2SF 2SF 2SF p=0
k =0 k =0 h=0

−1−q
2 SF 2 −1 2 −1
1   j2π SF−q
q SF−q
1 j2π [i−q ] k
−j2π [2 SF ] k p
+ SF e 2 SF ·e 2 SF = SF e 2 = 0.
2 2
k =2 SF −1−i p=0
h=0

Authorized licensed use limited to: ULAKBIM UASL - ANKARA UNIVERSITESI. Downloaded on January 09,2024 at 17:27:45 UTC from IEEE Xplore. Restrictions apply.
 1820 IEEE SIGNAL PROCESSING LETTERS, VOL. 24, NO. 12, DECEMBER 2017

As a conclusion for this section, we established the orthogonality

of the basis c(nTs + kT )|s(n T s )=i , i.e., that
 
c(nTs + kT )|s(n T s )=i , c(nTs + kT )|s(n T s )=q = 0

i = q, i, q ∈ {0 . . . 2SF − 1}. (9)

III. OPTIMUM DETECTION OF FSCM SIGNALS IN ADDITIVE

WHITE GAUSSIAN NOISE CHANNEL (AWGN) CHANNELS
Since we have equal energy signals and we suppose perfect
time and frequency synchronization as well as a source emit-
ting equally probable symbols, the optimum receiver for FSCM
signals in an AWGN channel can be derived easily a described
in [11, Sec. 6.1].
The received signal is

r(nTs + kT ) = c(nTs + kT ) + w(nTs + kT ) (10)
Fig. 1. Link level performance of FCSM and FSK modulations in the AWGN
where w(nTs + kT ) is a zero mean white gaussian noise, with (Flat) and frequency selective (Freq.Sel) channels.

σw2 (n T s +k T ) = σw2 independent of (nTs + kT ). The optimum
We can now rewrite the inner product in (11) as
demodulator consists of projecting r(nTs + kT ) onto the differ-  
ent signals c(nTs + kT )|s(n T s )=q , q = 0 . . . 2SF − 1 and choos- r(nTs + kT ), c(nTs + kT )|s(n T s )=p (15)
ing the signal c(nTs + kT )|s(n T s )=l such that the (square) mod-
ulus of the projection is maximum as the best estimate of the 2
SF
−1
k2 1
r(nTs + kT ) · e−j2π 2 SF · √ e−j2π pk 2 SF .
1

transmitted signal. This process is providing the best estimate = (16)

  2 SF
ŝ(nTs ) = l of the transmitted signal s(nTs ). k =0
d(n T s +k T )

From (16) we can see that the process of projecting

A. Computationally Efficient Implementation
the signal r(nTs + kT ) onto the signal basis element
To calculate the projection, we perform the usual computation c(nTs + kT )|s(n T s )=p consists of following two steps:
  1) multiplying the signal r(nTs + kT ) sample by sample by
r(nTs + kT ), c(nTs + kT )|s(n T s )=q k2
the signal e−j2π 2 SF (the so called down-chirp), obtaining
 k2
2 −1
the signal d(nTs + kTs ) = r(nTs + kT ) · e−j2π 2 SF ;
SF

= r(nTs + kT ) · c∗ (nTs + kT )|s(n T s )=q 2) taking the Discrete Fourier Transform of the vector
k =0 
d(nTs ), whose component d(nTs )k is [d(nTs )]k =
2−1
SF
d(nTs + kT ) and selecting the output of index p.
1
r(nTs + kT ) · √ e−j2π [(q +k ) m o d ]
k
= 2 SF 2 SF

k =0 2SF
IV. LINK LEVEL PERFORMANCE ANALYSIS
2
SF
−1
k2 1 In this section, we provide the results of some computer sim-
r(nTs + kT )e−j2π e−j2π [(q +k ) 2 SF −k 2 SF
] k
= 2 SF √ mod . ulations experiments where we compare the FSCM modulation
2 SF
k =0 ϕ(q ,k ) of cardinality 27 (i.e., with SF = 7) against an FSK modulation
(11) with the same cardinality. This means transmitting the waveform
We now turn our attention to the function 1 B
c(nTs + kT ) = √ ej2π s(n T s )k T 2 SF (17)
 −j2π [(q +k ) m o d 2SF
2 SF −k 2 SF
]
k
ϕ(q, k) = e − 1.
, q, k = 0, 1, . . . 2 SF

(12) instead of the one in (4) for the input symbol s(nTs ).
We can see that for q + k < 2SF , i.e., for k < 2SF − q, we have We have considered a flat frequency additive white gaus-
sian noise channel (AWGN) as well as a unit energy fre-
ϕ(q, k) = e−j2π [(q +k )−k ] 2 SF = e−j2π q ·k 2 SF .
k 1
(13) quency selective AWGN multipath
√ √ channel with impulse re-
Instead for q + k ≥ 2 , i.e., for k ≥ 2
SF SF
− q, we have sponse h(nT ) = 0.8δ(nT ) + 0.2δ(nT − T ). The results
are depicted in Fig. 1.
ϕ(q, k) = e−j2π [(q +k −2 )−k ] 2 SF
SF k
We can see that, while for the AWGN channel the perfor-
mance of the FCSM and FSK modulations are the same, for a
= e−j2π q ·k 2 SF · ej2π 2
1 SF 1
2 SF frequency selective channel FCSM outperforms FSK. The rea-
son is that any FCSM is sweeping the all frequency range and
= e−j2π q ·k 2 SF .
1
(14) kind of averaging the noise. Instead, the FSK signals falling the

Authorized licensed use limited to: ULAKBIM UASL - ANKARA UNIVERSITESI. Downloaded on January 09,2024 at 17:27:45 UTC from IEEE Xplore. Restrictions apply.
 VANGELISTA: FREQUENCY SHIFT CHIRP MODULATION: THE LORA MODULATION 1821

area of the channel where there is a larger attenuation generate [2] M. Centenaro, L. Vangelista, A. Zanella, and M. Zorzi, “Long-range com-
an effect on the bit error rate, which is more adverse that the munications in unlicensed bands: The rising stars in the IoT and smart
city scenarios,” IEEE Wireless Commun., vol. 23, no. 5, pp. 60–67, Oct.
beneficial effect of the FSK signals falling in the area of the 2016.
channel where there is actually an amplification. [3] J. P. Bardyn, T. Melly, O. Seller, and N. Sornin, “IoT: The era of LPWAN
is starting now,” in Proc. 42nd Eur. Solid-State Circuits Conf., 2016,
pp. 25–30.
V. CONCLUSION [4] F. Sforza, “Communication system,” U.S. Patent 8406275 B2, Mar. 26,
2013.
In this letter, we have – to the author’s knowledge – pro- [5] C. Goursaud and J.-M. Gorce, “Dedicated networks for IoT : PHY/MAC
vided the first complete mathematically rigorous description of state of the art and challenges,” EAI Endorsed Trans. Internet Things,
the LoRa modulation (i.e., FCSM) and demodulation processes, vol. 1, no. 1, pp. 1–11, 2015.
[6] B. Reynders, W. Meert, and S. Pollin, “Range and coexistence analy-
based on the (discrete time) signal theory and detection. We de- sis of long range unlicensed communication,” in Proc. 23rd Int. Conf.
rived the optimum demodulator and its efficient version. Finally, Telecomm., 2016, pp. 1–6.
we made a performance comparison of the LoRa Modulation [7] B. Reynders and S. Pollin, “Chirp spread spectrum as a modulation tech-
nique for long range communication,” in Proc. Symp. Commun. Veh. Tech-
and the FSK modulation in terms of uncoded bit error rate. We nol., 2016, pp. 1–5.
noted that, while the two modulations exhibit the same perfor- [8] M. R. Winkler, “Chirp signals for communications,” WESCON Conven-
mance in an AWGN channel, in a frequency selective channel tion Record Paper 14.2, 1962.
[9] C. E. Cook, “Linear FM signal formats for beacon and communication
the LoRa (i.e., FCSM) modulation performs better. systems,” IEEE Trans. Aerosp. Electron. Syst., vol. AES-10, no. 4, pp. 471–
The results presented in this letter will eventually enable a 478. Jul. 1974.
further assessment of the LoRa based networks much more [10] C. Gupta, T. Mumtaz, M. Zaman, and A. Papandreou-Suppappola, “Wide-
band chirp modulation for FH-CDMA wireless systems: Coherent and
rigorous than what has been done until now. non-coherent receiver structures,” in Proc. IEEE Int. Conf. Commun.,
2003, vol. 4, pp. 2455–2459.
REFERENCES [11] N. Benvenuto and G. Cherubini, Algorithms for Communications Systems
and Their Applications. Chichester, U.K.: Wiley, 2002.
[1] H. Wang and A. O. Fapojuwo, “A survey of enabling technologies of low [12] G. Cariolaro, Unified Signal Theory. Berlin, Germany: Springer, 2011.
power and long range machine-to-machine communications,” IEEE Com-
mun. Surv. Tuts., to be published doi: 10.1109/COMST.2017.2721379.

Authorized licensed use limited to: ULAKBIM UASL - ANKARA UNIVERSITESI. Downloaded on January 09,2024 at 17:27:45 UTC from IEEE Xplore. Restrictions apply.