Professional Documents
Culture Documents
process based on it
Andrea Burgosa) and Andre
s Santosb)
Departamento de Fısica, Universidad de Extremadura, 06006 Badajoz, Spain
(Received 28 January 2021; accepted 23 April 2021)
The Newcomb–Benford law, also known as the rst-digit law, gives the probability distribution
associated with the rst digit of a dataset so that, for example, the rst signicant digit has a probability
of 30.1% of being 1 and 4.58% of being 9. This law can be extended to the second and next signicant
digits. This article presents an introduction to the discovery of the law and its derivation from the scale
invariance property as well as some applications and examples. Additionally, a simple model of a
Markov process inspired by scale invariance is proposed. Within this model, it is proved that the
probability distribution irreversibly converges to the Newcomb–Benford law, in analogy to the
irreversible evolution toward equilibrium of physical systems in thermodynamics and statistical
mechanics. # 2021 Author(s). All article content, except where otherwise noted, is licensed under a Creative Commons
Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).
https://doi.org/10.1119/10.0004957
Fig. 1. Simon Newcomb (1835–1909). If the dataset fullls the NBL, then one has
852 Am. J. Phys., Vol. 89, No. 9, September 2021 A. Burgos and A. Santos 852
Fig. 4. Histograms showing the distribution of the rst digit for (from left to
right at each digit d) the NBL and the populations of the municipalities of
the provinces of Badajoz and Caceres, the region of Extremadura, and
Spain.
853 Am. J. Phys., Vol. 89, No. 9, September 2021 A. Burgos and A. Santos 853
converge toward the NBL. However, this is not the case. As
illustrated in Fig. 7 for the case k ¼ 2 and an initial dataset of
104 random numbers, the time-dependent distribution
fpd ðtÞg exhibits a quasiperiodic oscillatory pattern around
the NBL distribution without any apparent amplitude attenu-
ation. In fact, since 210 ¼ 1024 ’ 103 , the distribution nearly
returns to the uniform initial one at times t ¼ 10; 20; 30; ….
This behavior is reminiscent of the Poincare recurrence time
in mechanical systems and the associated Zermelo paradox
about irreversibility.51
In the transformation frn ðtÞg ! frn ðt þ 1Þ ¼ 2rn ðtÞg, the
rst-digit distribution changes, according to Fig. 3, as
854 Am. J. Phys., Vol. 89, No. 9, September 2021 A. Burgos and A. Santos 854
distributions of the rst digit, pd ðtÞ, and of the rst two digits, choose for ad the values given by Eqs. (5), the stationary
pd;d2 ðtÞ, of the dataset frn ðtÞg [see Eq. (4)]. As a conse- solution coincides exactly with the NBL, as could be
quence, (i) Eqs. (6) do not make a closed set and (ii) the expected. This will be the choice made in the rest of this
parameters ad depend on t. section.
At this point, we construct a simplied dynamical model To illustrate the irreversible evolution of pðtÞ toward p ,
in which the four unknown and time-dependent parameters we are going to consider two different initial conditions.
ad in Eqs. (6) are replaced by xed constants. In matrix First, we start from a uniform distribution, that is, pd ð0Þ ¼ 19.
notation, The result is shown in Fig. 8, where we see that the evolution
is oscillatory (see Appendix B for an explanation) and the
pðt þ 1Þ ¼ W pðtÞ; (7)
oscillations are rapidly damped to attain the stationary solu-
tion. As a second example, we take an NBL inverted initial
where pðtÞ ¼ ðp1 ðtÞ; p2 ðtÞ; …; p9 ðtÞÞ† is a column vector
distribution, that is, pd ð0Þ ¼ p10d , so that, initially, state 9 is
(† denotes transposition) and
the most probable and state 1 is the least probable. In this
0
0 0 0 0 1 1 1 1 1
1 second example, as shown in Fig. 9, the initial oscillations
B a are of greater amplitude but, as before, the stationary distri-
B 1 0 0 0 0 0 0 0 0 C
C bution is practically reached after a few iterations.
B
B 1 a1
C
B 0 0 0 0 0 0 0 0 C
C Comparison between Figs. 7 and 8 shows that the main dif-
B 0
B a2 0 0 0 0 0 0 0 C
C ference between our Markov model and the non-Markovian
W¼B 0
B
1 a2 0 0 0 0 0 0 0
C
C evolution of fpd ðtÞg in a real dataset is that the oscillation
B C amplitudes are attenuated in the model and not in the original
B 0
B 0 a3 0 0 0 0 0 0 C
C framework.
B 0
B 0 1 a2 0 0 0 0 0 0 C
C It seems convenient to characterize the evolution of the set
@ 0
B
0 0 a4 0 0 0 0 0
C
A of probabilities fpd ðtÞg obeying the Markov process [Eq.
0 0 0 1 a4 0 0 0 0 0 (7)] toward the attractor fpd g by means of a single parameter
that, in addition, evolves monotonically, thus representing
(8) the irreversibility of evolution. It is expected that these prop-
is a xed square matrix. Equation (7) ts the canonical
erties are veried by the Kullback–Leibler divergence,53
description of Markov chains,52 where W is the so-called
which in our case is dened as
transition matrix, and fad g correspond to transition probabil-
ities. In this way, we may forget about the meaning of
Fig. 8. Evolution of the ratio pd ðtÞ=pd (where a vertical shift d 1 has been Fig. 9. Evolution of the ratio pd ðtÞ=pd (where a vertical shift d 1 has been
applied for better clarity), when starting from a uniform initial distribution applied for better clarity), when starting from an inverted initial distribution
pd ð0Þ ¼ 19, according to our Markov-chain model, Eq. (7). pd ð0Þ ¼ p10d , according to our Markov-chain model, Eq. (7).
855 Am. J. Phys., Vol. 89, No. 9, September 2021 A. Burgos and A. Santos 855
system. In both cases, a statistical description is constructed
by introducing the adequate probability distribution: the
velocity distribution function (gas) and the probability of
the internal state d (Markov chain); while f ð~ v; tÞ is continu-
ous in both ~v and t, pd ðtÞ is discrete in d and t. The evolu-
tion equation for the probability distribution function is the
Boltzmann equation (gas, under the molecular chaos
ansatz56) and Eq. (7) (Markov chain). The equilibrium state
is characterized by the Maxwell–Boltzmann distribution57
vÞ (gas) and the NBL distribution pd (Markov chain).
fMB ð~
In both classes of systems, the nonequilibrium entropy
functional S(t) can be dened, within a constant, as the
opposite of the Kullback–Leibler divergence53,58 of the
equilibrium distribution from the time-dependent one.
Boltzmann’s celebrated H-theorem56,59 (gas) and the result
presented in Eq. (10) (Markov chain) show that the entropy
S(t) never decreases, irreversibly evolving toward its equi-
librium value.
Fig. 10. Evolution of the Kullback–Leibler divergence DKL ðtÞ (in logarith-
V. CONCLUDING REMARKS
mic scale), starting from the uniform initial distribution pd ð0Þ ¼ 19 and from
the inverted initial distribution pd ð0Þ ¼ p10d , according to our Markov- One of the main goals of this article was to show that, con-
chain model, Eq. (7). The dashed line is proportional to ja2;3 j2t (see
Appendix C).
trary to what might be initially thought, the rst signicant
digit d of a dataset extracted from measurements or observa-
tions of the real world is not evenly distributed among the
nine possible values, but typically the frequency is higher for
Figure 10 shows the evolution of DKL ðtÞ for the same ini- d ¼ 1 and decreases as d increases. The NBL, Eq. (1), gives a
tial conditions as in Figs. 8 and 9. Both cases conrm the mathematical expression to this empirical fact, although it
desired monotonic evolution of DKL ðtÞ. Also, the asymptotic does not always need to be rigorously veried due to statisti-
decay DKL ðtÞ ! 0 occurs essentially exponentially with a cal uctuations (as happens with populations in Fig. 4 and
Table II. Analogy between a classical dilute gas (as a prototypical physical system) and a system described by the Markov chain, Eq. (7). In the expression of
the Maxwell–Boltzmann distribution fMB ð~ vÞ, m is the mass of a molecule, T is the gas temperature, and kB is the Boltzmann constant. In the Boltzmann equa-
vjf ðtÞ; f ðtÞ is the collision operator.
tion, J½~
856 Am. J. Phys., Vol. 89, No. 9, September 2021 A. Burgos and A. Santos 856
exactly solved, the solution showing an irreversible relaxa- sides of the equation with respect to k and then taking k ¼ 1,
tion toward the NBL probability distribution. Moreover, we we easily obtain xP0x ðxÞ ¼ Px ðxÞ, which, according to
have proved that the associated (relative) entropy monotoni- Euler’s theorem, implies that Px ðxÞ is a homogeneous func-
cally increases, in analogy with the second law of thermody- tion of degree 1, that is, Px ðxÞ / x1 . The constant of pro-
namics in physical systems. portionality is obtained from the normalization condition,
Until the 1970s (which is when scientic pocket calcula- nally yielding
tors began to be used), physicists used tables of logarithms
(or their application in slide rules) for small everyday scien- x1
Px ðxÞ ¼ ; 0 x < 10: (A3)
tic calculations. Those calculations are nowadays per- ln 10
formed on pocket calculators, cellular phones, or personal
This is the unique distribution of signicands that is invariant
computers with a wide variety of existing mathematical pro- under a change of scale. From Eq. (A3), and applying Eqs.
grams. Since the data that are manipulated in physics are (A1) and (A2), it is straightforward to obtain NBL, Eq. (1),
extracted from “real” situations, such as experiments, mod- and its generalization, Eq. (2). As a consistency test of Eq.
els, physical constants, …, we can conclude, as a tribute to (2), it is easy to check that
Newcomb and Benford and their logarithmic tables, that key-
board button 1 will be the one that presents the greatest wear 9
X
and tear, while that of 9 will be the least used. pd1 ;d2 ;…;dm1 ¼ pd1 ;d2 ;…;dm
dm ¼0
2 !1 3
ACKNOWLEDGMENTS m1
X
¼ log10 41 þ di 10m1i 5: (A4)
A.S. acknowledges nancial support from Grant No.
i¼1
FIS2016-76359-P/AEI/10.13039/501100011033 and the
Junta de Extremadura (Spain) through Grant No. GR18079, It is interesting to note that the inverse law for the signi-
both partially nanced by Fondo Europeo de Desarrollo cand, Eq. (A3), implies a uniform law for the mantissa (and
Regional funds. vice versa). Let Pl ðlÞdl be the probability that the mantissa
APPENDIX A: DERIVATION OF THE NBL AND lies between l and l þ dl. Since Pl ðlÞdl ¼ Px ðxÞdx and
SOME GENERALIZATIONS dl ¼ ðx1 = ln 10Þdx, Eq. (A3) gives us Pl ðlÞ ¼ 1. In
Newcomb’s words,1 “The law of probability of the occur-
Consider a long list of records frn g that, without loss of rence of numbers is such that all mantissæ of their logarithms
857 Am. J. Phys., Vol. 89, No. 9, September 2021 A. Burgos and A. Santos 857
is the unique stationary solution. From the normalization
1
pd ¼ logb 1þ ; d ¼ 1; 2; …; b 1: (A6) condition, one has p9 ¼ a1 a2 ð1 a4 Þ=ð3 þ a1 a2 Þ. Note that,
d
as seen from Eqs. (B3), the initial values pII ð0Þ do not inu-
ence the evolution of either pI ðtÞ or pII ðtÞ.
APPENDIX B: EXACT SOLUTION OF THE The stationary solution will be an attractor if
limt!1 pI ðtÞ ¼ pI and limt!1 pII ðtÞ ¼ pII for any initial con-
MARKOV-CHAIN MODEL
dition pI ð0Þ. According to Eqs. (B3), this implies
P9
Note rst that Eqs. (6) verify the normalization condition
P9 limt!1 At ¼ 0.
d¼1 pd ðt þ 1Þ ¼ d¼1 pd ðtÞ ¼ 1. Therefore, only eight of
To check the above condition, let us obtain the four eigen-
the probabilities fpd ; d ¼ 1; 2; …; 9g are linearly indepen- values fai ; i ¼ 0; 1; 2; 3g of A. It is easy to see that the char-
dent, so we can eliminate one of them. If we choose p9 ¼ 1 acteristic equation is aða1 a2 þ a þ a2 þ a3 Þ ¼ 0. Therefore,
8d¼1 pd , Eq. (6a) gives us p1 ðt þ 1Þ ¼ 1 p1 ðtÞ p2 ðtÞ
P the eigenvalues are a0 ¼ 0 and
p3 ðtÞ p4 ðtÞ. Although it is not strictly necessary, it is
1 2
mathematically convenient to split the column vector pðtÞ a1 ¼ 1þ b ; (B5a)
into the subsets pI ðtÞ ðp1 ðtÞ; p2 ðtÞ; p3 ðtÞ; p4 ðtÞÞ† , pII ðtÞ 3 b
ðp5 ðtÞ; p6 ðtÞ; p7 ðtÞ; p8 ðtÞÞ† , and p9 ðtÞ. As a consequence, p p !
the matrix equation (7) can be decomposed into 1 16ı 3 17ı 3
a2;3 ¼ 1 þ b ; (B5b)
3 b 2
pI ðt þ 1Þ ¼ q þ A pI ðtÞ; pII ðt þ 1Þ ¼ B pI ðtÞ; (B1)
858 Am. J. Phys., Vol. 89, No. 9, September 2021 A. Burgos and A. Santos 858
from Table I. On the other hand, the most natural choice is DDKL ðt þ 1Þ DKL ðt þ 1Þ DKL ðtÞ
provided by Eqs. (5), in which case the stationary solution 9
coincides exactly with the NBL.
X
9
pd ðtÞ pd0
d0 ¼5
X
¼ pd ðtÞ ln 9
: (C6)
APPENDIX C: PROPERTIES OF THE d¼5
X
pd pd0 ðtÞ
KULLBACK–LEIBLER DIVERGENCE d 0 ¼5
IN THE MARKOV MODEL
Given the stationary values fpd ; d ¼ 5; …; 9g, the difference
The Kullback–Leibler divergence is dened by Eq. (9). In
DDKL ðt þ 1Þ is a function of the ve probabilities fpd ðtÞ; d
order to analyze its asymptotic decay, let us consider times
¼ 5; …; 9g. To nd the maximum value of DDKL ðt þ 1Þ, we
that are long enough so that the deviations dpd ðtÞ
take the derivative
pd ðtÞ pd can be considered small. In this regime, we can
expand Eq. (9) in a power series and retain the dominant 9
X
terms. The result is pd ðtÞ pd0
@DDKL ðt þ 1Þ d0 ¼5
1X 9
½dpd ðtÞ2 ¼ ln : (C7)
DKL ðtÞ : (C1) @pd ðtÞ 9
X
2 d¼1 pd pd pd0 ðtÞ
d 0 ¼5
On the other hand, for times long enough, ja1 jt ja2;3 jt (note
The unique solution to the extremum conditions @DDKL ðt
that ja1 j ¼ 0:4261 and ja2;3 j ¼ 0:8692), so that, according to
þ1Þ=@pd ðtÞ ¼ 0 is pd ðtÞ ¼ cpd (d ¼ 5; …; 9), where 0 < c
Eqs. (B3), dpd ðtÞ ja2;3 jt . Thus,
< 1= 9d¼5 pd is arbitrary. In such a case, DDKL ðt þ 1Þ ¼ 0.
P
DKL ðtÞ ja2;3 j2t ¼ 102t log10 ja2;3 j : (C2) To see that this is actually a maximum value, suppose, for
instance, that pd ðtÞ ¼ 0 except if d ¼ d0 (with d0 ¼ 5; …; 9).
This asymptotic behavior is represented in Fig. 10. In that case, it is easy to nd DDKL ðt þ 1Þ
¼ pd0 ðtÞ ln ðpd0 = 9d¼5 pd Þ < 0. We then conclude that
P
Finally, let us prove Eq. (10). According to Eq. (9),
DDKL ðt þ 1Þ 0, i.e., we obtain Eq. (10), the equality hold-
p1 ðt þ 1Þ ing only if pd ðtÞ ¼ cpd (d ¼ 5; …; 9). Note that, even though
DKL ðt þ 1Þ ¼ p1 ðt þ 1Þ ln
p1 DDKL ðt þ 1Þ ¼ 0 if pd ðtÞ=pd ¼ constant for d ¼ 5; …; 9, this
859 Am. J. Phys., Vol. 89, No. 9, September 2021 A. Burgos and A. Santos 859
pages of logarithmic tables,” J. Assoc. Inf. Sci. Technol. 69, 349–358 Curry, and A. M. Dougherty, “Stochastic aspects of one-dimensional dis-
(2018); D. S. Lemons, “Thermodynamics of Benford’s rst digit law,” crete dynamical systems: Benford’s law,” Phys. Rev. E 64, 026222
Am. J. Phys. 87, 787–790 (2019). (2001); A. Berger, L. A. Bunimovich, and T. P. Hill, “One-dimensional
8
B. Buck, A. C. Merchant, and S. M. Perez, “An illustration of Benford’s dynamical systems and Benford’s law,” Trans. Am. Math. Soc. 357,
rst digit law using alpha decay half lives,” Eur. J. Phys. 14, 59–63 197–219 (2004).
24
(1993). Q. Li, Z. Fu, and N. Yuan, “Beyond Benford’s law: Distinguishing noise
9
T. P. Hill, “The rst digit phenomenon: A century-old observation about from chaos,” PLoS One 10, e0129161 (2015).
25
an unexpected pattern in many numerical tables applies to the stock mar- L. Shao and B.-Q. Ma, “The signicant digit law in statistical physics,”
ket, census statistics and accounting data,” Am. Sci. 86, 358–363 (1998). Physica A 389, 3109–3116 (2010).
10 26
R. S. Pinkham, “On the distribution of rst signicant digits,” Ann. Math. J. M. Campanario and M. A. Coslado, “Benford’s law and citations,
Stat. 32, 1223–1230 (1961); B. J. Flehinger, “On the probability that a ran- articles and impact factors of scientic journals,” Scientometrics 88,
dom integer has initial digit A,” Am. Math. Mon. 73, 1056–1061 (1966); 421–432 (2011); A. D. Alves, H. H. Yanasse, and N. Y. Soma, “Benford’s
R. A. Raimi, “On the distribution of rst signicant gures,” ibid. 76, law and articles of scientic journals: comparison of JCR and Scopus
342–348 (1969); “The rst digit problem,” 83, 521–538 (1976); P. data,” ibid. 98, 173–184 (2014).
Diaconis, “The distribution of leading digits and uniform distribution mod 27
M. J. Nigrini and S. J. Miller, “Data diagnostics using second-order tests
1,” Ann. Probab. 5, 72–81 (1977); K. Nagasaka, “On Benford’s law,” of Benford’s law,” Auditing 28, 305–324 (2009).
Ann. Inst. Stat. Math. 36, 337–352 (1984); T. P. Hill, “A statistical deriva- 28
W. K. T. Cho and B. J. Gaines, “Breaking the (Benford) law. Statistical
tion for the signicant-digit law,” Stat. Sci. 10, 354–363 (1995); L. M. fraud detection in campaign nance,” Am. Stat. 61, 218–223 (2007); B. F.
Leemis, B. W. Schmeiser, and D. L. Evans, “Survival distributions satisfy- Roukema, “A rst-digit anomaly in the 2009 Iranian presidential election,”
ing Benford’s law,” Am. Stat. 54, 236–241 (2000); M. Cong, C. Li, and J. Appl. Stat. 41, 164–199 (2014); D. Gamermann and F. L. Antunes,
B.-Q. Ma, “First digit law from Laplace transform,” Phys. Lett. A 383, “Statistical analysis of Brazilian electoral campaigns via Benford’s law,”
1836–1844 (2019); M. Cong and B.-Q. Ma, “A proof of rst digit law Physica A 496, 171–188 (2018).
from Laplace transform,” Chin. Phys. Lett. 36, 070201 (2019); A. Volčič, 29
A. Diekmann, “Not the rst digit! Using Benford’s law to detect fraudu-
“Uniform distribution, Benford’s law and scale-invariance,” Boll. Unione lent scientic data,” J. Appl. Stat. 34, 321–329 (2007).
Mat. Ital. 13, 539–543 (2020); A. Berger and T. P. Hill, “The mathematics 30
J. Shi, M. Ausloos, and T. Zhu, “Benford’s law rst signicant digit and
of Benford’s law: a primer,” Stat. Methods Appl. (published online). distribution distances for testing the reliability of nancial reports in devel-
11
T. P. Hill, “Base-invariance implies Benford’s law,” Proc. Am. Math. Soc. oping countries,” Physica A 492, 878–888 (2018); M. Ausloos, A.
123, 887–895 (1995). Eskandary, P. Kaur, and G. Dhesi, “Evidence for gross domestic product
12
A. Berger and T. P. Hill, “A basic theory of Benford’s law,” Probab. Surv. growth time delay dependence over foreign direct investment. A time-lag
8, 1–126 (2011).
13 dependent correlation study,” ibid. 527, 121181 (2019).
E. W. Weisstein, “Benford’s law,” <https://mathworld.wolfram.com/ 31
L. Pietronero, E. Tosatti, V. Tosatti, and A. Vespignani, “Explaining the
BenfordsLaw.html>.
14 uneven distribution of numbers in nature: the laws of Benford and Zipf,”
D. C. Hoyle, M. Rattray, R. Jupp, and A. Brass, “Making sense of microar-
Physica A 293, 297–304 (2001).
ray data distributions,” Bioinformatics 18, 576–584 (2002). 32
15 M. Miranda-Zanetti, F. Delbianco, and F. Tohme, “Tampering with ina-
860 Am. J. Phys., Vol. 89, No. 9, September 2021 A. Burgos and A. Santos 860
44
K.-B. Lee, S. Han, and Y. Jeong, “COVID-19, attening the curve, and Gases in Shear Flows: Nonlinear Transport, Fundamental Theories of
Benford’s law,” Physica A 559, 125090 (2020); A. P. Kennedy and S. C. Physics (Springer, Netherlands, 2003); H. Karabulut, “Direct simulation
P. Yam, “On the authenticity of COVID-19 case gures,” PLoS One 15, for a homogeneous gas,” Am. J. Phys. 75, 62–66 (2007); A. D. Boozer,
e0243123 (2020). “Thermodynamic time asymmetry and the Boltzmann equation,” ibid. 83,
45
“Testing Benford’s law,” <https://testingbenfordslaw.com>. 223–230 (2016).
46 57
“Instituto Nacional de Estadıstica,” <https://www.ine.es/index.htm>. J. Hurley, “Equilibrium solution of the Boltzmann equation,” Am. J. Phys.
47
“The brightest stars,” <http://www.atlasoftheuniverse.com/stars.html>. 33, 242–243 (1965); P. A. Mello and T. A. Brody, “A different proof of
48
“Sunspot number,” <http://sidc.be/silso/datales>. the Maxwell-Boltzmann distribution,” ibid. 40, 1239–1345 (1972); R.
49
“Corteel,” <https://corteel.com/es/es/mujer?srule=price-high-to-low>. Lopez-Ruiz and X. Calbet, “Derivation of the Maxwellian distribution
50
“Hipercor,” <https://www.hipercor.es/supermercado/alimentacion/>. from the microcanonical ensemble,” ibid. 75, 752–753 (2007); A.
51
V. S. Steckline, “Zermelo, Boltzmann, and the recurrence paradox,” Am. Walstad, “On deriving the Maxwellian velocity distribution,” ibid. 81,
J. Phys. 51, 894–897 (1982). 555–557 (2013); R. E. Robson, T. J. Mehrling, and J. Osterhoff, “Great
52
Y. K. Leong, “Mechanical model for 2–state Markov chains,” Am. J. moments in kinetic theory: 150 years of Maxwell’s (other) equations,”
Phys. 52, 749–753 (1984). Eur. J. Phys. 38, 065103 (2017); F. Rivadulla, “Alternative derivation of
53
S. Kullback and R. A. Leibler, “On information and sufciency,” Ann. the Maxwell distribution of speeds,” J. Chem. Educ. 96, 2063–2065
Math. Stat. 22, 79–86 (1951). (2019).
54 58
J. Machta, “Entropy, information, and computation,” Am. J. Phys. 67, S. Tiwary, “Time evolution of entropy, in various scenarios,” Eur. J. Phys.
1074–1077 (1999). 41, 025101 (2020).
55 59
A. Ben-Naim, Entropy Demystied: The Second Law Reduced to Plain J. Pojman, “Boltzmann’s H theorem applied to simulations of polymer
Common Sense (World Scientic, Hackensack, NJ, 2008). interchange reactions,” J. Chem. Educ. 67, 200–202 (1990); A. D. Boozer,
56
S. Chapman and T. G. Cowling, The Mathematical Theory of Non- “Boltzmann’s H-theorem and the assumption of molecular chaos,” Eur. J.
Uniform Gases, 3rd ed. (Cambridge U. P., Cambridge, UK, 1970); C. Phys. 32, 1391–1403 (2011); K. Rȩbilas, “Origin of the thermodynamic
Cercignani, The Boltzmann Equation and Its Applications (Springer- time arrow demonstrated in a realistic statistical system,” Am. J. Phys. 80,
Verlag, New York, 1988); V. Garz o and A. Santos, Kinetic Theory of 700–707 (2012).
861 Am. J. Phys., Vol. 89, No. 9, September 2021 A. Burgos and A. Santos 861