Laws of Population Growth

Abstract

An important issue in the study of cities is defining a metropolitan area, because different definitions affect conclusions regarding the statistical distribution of urban activity. A commonly employed method of defining a metropolitan area is the Metropolitan Statistical Areas (MSAs), based on rules attempting to capture the notion of city as a functional economic region, and it is performed by using experience. The construction of MSAs is a time-consuming process and is typically done only for a subset (a few hundreds) of the most highly populated cities. Here, we introduce a method to designate metropolitan areas, denoted "City Clustering Algorithm" (CCA). The CCA is based on spatial distributions of the population at a fine geographic scale, defining a city beyond the scope of its administrative boundaries. We use the CCA to examine Gibrat's law of proportional growth, which postulates that the mean and standard deviation of the growth rate of cities are constant, independent of city size. We find that the mean growth rate of a cluster by utilizing the CCA exhibits deviations from Gibrat's law, and that the standard deviation decreases as a power law with respect to the city size. The CCA allows for the study of the underlying process leading to these deviations, which are shown to arise from the existence of long-range spatial correlations in population growth. These results have sociopolitical implications, for example, for the location of new economic development in cities of varied size.

Full text

arXiv:0808.2202v2 [physics.soc-ph] 10 Sep 2009
Laws of Population Growth
Hern´an D. Rozenfeld1, Diego Rybski1, Jos´e S. Andrade Jr.2,
Michael Batty3, H. Eugene Stanley4, and Hern´an A. Makse1,2
1Levich Institute and Physics Department,
City College of New York, New York, NY 10031, USA
2Departamento de F´ısica, Universidade Federal
do Cear´a, 60451-970 Fortaleza, Cear´a, Brazil
3Centre for Advanced Spatial Analysis, University College London,
1-19 Torrington Place, London WC1E 6BT, UK
4Center for Polymer Studies and Physics Department,
Boston University, Boston, MA 02215, USA
Abstract
An important issue in the study of cities is defining a metropolitan area, as different definitions
affect conclusions regarding the statistical distribution of urban activity. A commonly employed
method of defining a metropolitan area is the Metropolitan Statistical Areas (MSAs), based on
rules attempting to capture the notion of city as a functional economic region, and is performed
using experience. The construction of MSAs is a time-consuming process and is typically done only
for a subset (a few hundreds) of the most highly populated cities. Here, we introduce a new method
to designate metropolitan areas, denoted “City Clustering Algorithm” (CCA). The CCA is based
on spatial distributions of the population at a fine geographic scale, defining a city beyond the
scope of its administrative boundaries. We use the CCA to examine Gibrat’s law of proportional
growth, which postulates that the mean and standard deviation of the growth rate of cities are
constant, independent of city size. We find that the mean growth rate of a cluster utilizing the CCA
exhibits deviations from Gibrat’s law, and that the standard deviation decreases as a power-law
with respect to the city size. The CCA allows for the study of the underlying process leading to
these deviations, which are shown to arise from the existence of long-range spatial correlations in
population growth. These results have socio-political implications, for example for the location of
new economic development in cities of varied size.
1

I. INTRODUCTION
In recent years there has been considerable work on how to define cities and how the
different definitions affect the statistical distribution of urban activity [1, 2]. This is a
long standing problem in spatial analysis of aggregated data sources, referred to as the
‘modifiable areal unit problem’ or the ‘ecological fallacy’ [3, 4], where different definitions of
spatial units based on administrative or governmental boundaries, give rise to inconsistent
conclusions with respect to explanations and interpretations of data at different scales. The
conventional method of defining human agglomerations is through the MSAs [1, 2, 5, 6, 7],
which is subject to socio-economical factors. The MSA has been of indubitable importance
for the analysis of population growth, and is constructed manually case-by-case based on
subjective judgment (MSAs are defined starting from a highly populated central area and
adding its surrounding counties if they have social or economical ties).
In this report, we propose a new way to measure the extent of human agglomerations
based on clustering techniques using a fine geographical grid, covering both urban and
rural areas. In this view, “cities” represent clusters of population, i.e., adjacent populated
geographical spaces. Our algorithm, the “city clustering algorithm” (CCA), allows for an
automated and systematic way of building population clusters based on the geographical
location of people. The CCA has one parameter (the cell size) that is useful for the study
of human agglomerations at different length scales, similar to the level of aggregation in
the context of social sciences. We show that the CCA allows for the study of the origin
of statistical properties of population growth. We use the CCA to analyze the postulates
of Gibrat’s law of proportional growth applied to cities, which assumes that the mean and
standard deviation of the growth rates of cities are constant. We show that population
growth at a fine geographical scale for different urban and regional systems at country
and continental levels (Great Britain, the USA, and Africa) deviates from Gibrat’s law.
We find that the mean and standard deviation of population growth rates decrease with
population size, in some cases following a power-law behavior. We argue that the underlying
demographic process leading to the deviations from Gibrat’s law can be modeled from the
existence of long-range spatial correlations in the growth of the population, which may arise
from the concept that “development attracts further development.” These results have
implications for social policies, such as those pertaining to the location of new economic
2

development in cities of different sizes. The present results imply that, on average, the
greatest growth rate occurs in the smallest places where there is the greatest risk of failure
(larger fluctuations). A corollary is that the safest growth occurs in the largest places having
less likelihood for rapid growth.
The analyzed data consist of the number of inhabitants, ni(t), in each cell iof a fine
geographical grid at a given time, t. The cell size varies for each data set used in this study.
We consider three different geographic scales: on the smallest scale, the area of study is Great
Britain (GB: England, Scotland and Wales), a highly urbanized country with population
of 58.7 million in 2007, and an area of 0.23 million km2. The grid is composed of 5.75
million cells of 200m-by-200m [8]. At the intermediate scale, we study the USA (continental
USA without Alaska), a single country nearly continental in scale, with a population of
303 million in 2007, and an area of 7.44 million km2. The grid contains 7.44 million cells
of approximately 1km-by-1km obtained from the US Census Bureau [9]. The datasets of
GB and USA are populated-places datasets, with population counts defined at points in a
grid. Since there could be some distortions in the true residential population involved at
the finest grid resolution, we perform our analysis by investigating the statistical properties
as a function of the grid size by coarse-graining the data as explained in Section IV A. At
the largest scale, we analyze the continent of Africa, composed of 53 countries with a total
population of 933 million in 2007, and an area of 30.34 million km2. These data are gridded
with less resolution by 0.50 million cells of approximately 7.74km-by-7.74km [10]. More
detailed information about these datasets is found in Section IV A (all the datasets studied
in this paper are available at http://lev.ccny.cuny.edu/∼hmakse/cities/city data.zip).
II. RESULTS
Figure 1A illustrates operation of the CCA. In order to identify urban clusters, we require
connected cells to have nonzero population. We start by selecting an arbitrary populated
cell (final results are independent of the choice of the initial cell). Iteratively, we then grow a
cluster by adding nearest neighbors of the boundary cells with a population strictly greater
than zero, until all neighbors of the boundary are unpopulated. We repeat this process until
all populated cells have been assigned to a cluster. This technique was introduced to model
forest fire dynamics [11] and is termed the “burning algorithm,” since one can think of each
3

populated cell as a burning tree.
The population Si(t) of cluster iat time tis the sum of the populations n(i)
j(t)
of each cell jwithin it, Si(t) = PNi
j=1 n(i)
j(t), where Niis the number of cells
in the cluster. Results of the CCA are shown in Fig. 1B, representing the urban
cluster surrounding the City of London (red cluster overlaying a satellite image, see
http://lev.ccny.cuny.edu/∼hmakse/cities/london.gif for an animated image of Fig. 1B). Fig-
ure 1C depicts all the clusters of GB, indicating the large variability in their population and
size.
A feature of the CCA is that it allows the analysis of the population clusters at different
length scales by coarse-graining the grid and applying the CCA to the coarse-grained dataset
(see Section IV A for details on coarse-graining the data). At larger scales, disconnected
areas around the edge of a cluster could be added into the cluster. This is justified when,
for example, a town is divided by a wide highway or a river.
Tables I and II in Supporting Information (SI) Section I. show a detailed comparison
between the urban clusters obtained with the CCA applied to the USA in 1990, and the
results obtained from the analysis of MSAs from the US Census Bureau used in previous
studies of population growth [5, 6, 7]. We observe that the MSAs considered in Ref. [5] are
similar to the clusters obtained with the CCA with a cell size of 4km-by-4km or 8km-by-
8km. In particular, the population sizes of the clusters have the same order of magnitude
as the MSAs. On the other hand, for large cities the MSAs from the data of Ref. [6] seem
to be mostly comparable to our results for cell sizes of 2km-by-2km or 4km-by-4km.
Use of the CCA permits a systematic study of cluster dynamics. For instance, clus-
ters may expand or contract, merge or split between two considered times as illustrated
in Fig. 2. We quantify these processes by measuring the probability distribution of the
temporal changes in the clusters for the data of GB. We find that when the cell size is
2.2km-by-2.2km, 84% of the clusters evolve from 1981 to 1991 following the three first cases
presented in Fig. 2 (no change, expansion or reduction), 6% of the clusters merge from two
clusters into one in 1991, and 3% of the clusters split into two clusters.
Next, we apply the CCA to study the dynamics of population growth by investigating
Gibrat’s law, which postulates that the mean and standard deviation of growth rates are
constant [1, 2, 5, 7, 12]. The conventional method [1, 2, 7] is to assume that the populations
4

of a given city or cluster i, at times t0and t1> t0, are related by
S1=R(S0)S0,(1)
where S0≡Si(t0) = PNi
jn(i)
j(t0) and S1≡Si(t1) = PNi
jn(i)
j(t1) are the initial and final
populations of cluster i, respectively, and R(S0) is the positive growth factor which varies
from cluster to cluster. Following the literature in population dynamics [1, 2, 5, 7], we
define the population growth rate of a cluster as r(S0)≡ln R(S0) = ln(S1/S0), and study
the dependence of the mean value of the growth rate, hr(S0)i, and the standard deviation,
σ(S0) = phr(S0)2i − hr(S0)i2, on the initial population, S0. The averages hr(S0)iand σ(S0)
are calculated applying nonparametric techniques [13, 14] (see Section IV B for details). To
obtain the population growth rate of clusters we take into account that not all clusters
occupy the same area between t0and t1according to the cases discussed in Fig. 2. The
figure shows how to calculate the growth rate r(S0) in each case.
We analyze the population growth in the USA from t0= 1990 to t1= 2000 [9]. We apply
the CCA to identify the clusters in the data of 1990 and calculate their growth rates by
comparing them to the population of the same clusters in 2000 when the data are gridded
with a cell size of 2000m by 2000m. We calculate the annual growth rates by dividing rby
the time interval t1−t0.
Figure 3A shows a nonparametric regression with bootstrapped 95% confidence bands [13,
14] of the growth rate of the USA, hr(S0)i(see Section IV B for details). We find that the
growth rate diminishes from hr(S0)i ≈ 0.012 ±0.004 (error includes the confidence bands)
for populations below 104inhabitants to hr(S0)i ≈ 0.002 ±0.002 for the largest populations
around S0≈107. We may argue that the mean growth rate deviates from Gibrat’s law
beyond the confidence bands. While it is difficult to fit the data to a single function for the
entire range, the data show a decrease with S0approximately following a power-law in the
tail for populations larger than 104. An attempt to fit the data with a power-law yields the
following scaling in the tail:
hr(S0)i ∼ S−α
0,(2)
where αis the mean growth exponent, that takes a value αUSA = 0.28 ±0.08 from Ordinary
Least Squares (OLS) analysis [15] (see Section IV B for details on OLS and on the estimation
of the exponent error).
5

Figure 3B shows the dependence of the standard deviation σ(S0) on the initial population
S0. On average, fluctuations in the growth rate of large cities are smaller than for small
cities in contrast to Gibrat’s law. This result can be approximated over many orders of
magnitude by the power-law,
σ(S0)∼S−β
0,(3)
where βis the standard deviation exponent. We carry out an OLS regression analysis and
find that βUSA = 0.20 ±0.06. The presence of a power-law implies that fluctuations in the
growth process are statistically self-similar at different scales, for populations ranging from
∼1000 to ∼10 million according to Fig. 3B.
Figure 4 shows the analysis of the growth rate of the population clusters of GB from
gridded databases [8] with a cell size of 2.2km-by-2.2km at t0= 1981 and t1= 1991. The
average growth rate depicted in Fig. 4A comprises large fluctuations as a function of S0,
especially for smaller populations. However, a slight decrease with population seems evident
from rates around hri ≈ 0.008±0.001 with S0≈104dropping to zero or even negative values
for the largest populations, S0≈106. We find that 3556 clusters with population around
S0= 103exhibit negative growth rates as well. Thus, the mean rates are plotted on a
semi-logarithmic scale in Fig. 4A. When considering intermediate populations ranging from
S0= 3000 to S0= 3 ×105, the data seem to be following approximately a power-law
with αGB = 0.17 ±0.05 from OLS regression analysis, as shown in the inset of Fig. 4A.
Figure 4B shows the standard deviation for GB, σ(S0), exhibiting deviations from Gibrat’s
law having a tendency to decrease with population according to Eq. (3) and a standard
deviation exponent, βGB = 0.27 ±0.04, obtained with OLS technique.
The CCA allows for a study of the growth rates as a function of the scale of observation, by
changing the size of the grid. We find (SI Section II.) that the data for GB are approximately
invariant under coarse-graining the grid at different levels for both the mean and standard
deviation. When the data of the USA are aggregated spatially from cell size 2000m to
8000m, the scaling of the mean rates crosses-over to a flat behavior closer to Gibrat’s law.
At the scale of 8000m the mean is approximately constant (with fluctuations). However,
we find that, at this scale, all cities in the northeastern the USA spanning from Boston to
Washington D.C. form a single cluster. Despite these differences, the scaling of the standard
deviation for the USA holds approximately invariant even up to the large scale of observation
of 8000m.
6

Next, we analyze the population growth in Africa during the period 1960 to 1990 [10].
In this case the population data are based on a larger cell size, so we evaluate the data
cell by cell (without the application of the CCA). Despite the differences in the economic
and urban development of Africa, Great Britain and the USA, we find that the mean and
standard deviation of the growth rate in Africa display similar scaling as found for the USA
and GB. In Fig. 5A we show the results for the growth rate in Africa when the grid is
coarse-grained with a cell size of 77km-by-77km. We find a decrease of the growth rate from
hr(S0)i ≈ 0.1 to hr(S0)i ≈ 0.01 between populations S0≈103and S0≈106, respectively.
All populations have positive growth rates. A log-log plot of the mean rates shown in
Fig. 5A reveals a power-law scaling hr(S0)i ∼ S−αAf
0, with αAf = 0.21 ±0.05 from OLS
regression analysis. The standard deviation (Fig. 5B) satisfies Eq. (3) with a standard
deviation exponent βAf = 0.19 ±0.04. The CCA allows for a study of the origin of the
observed behavior of the growth rates by examining the dynamics and spatial correlations
of the population of cells. To this end, we first generate a surrogate dataset that consists of
shuffling two randomly chosen populated cells, n(i)
j(t0) and n(i)
k(t0), at time t0. This swapping
process preserves the probability distribution of n(i)
j, but destroys any spatial correlations
among the population cells. Figure 4C shows the results of the randomization of the GB
dataset, indicating power-law scaling in the tail of σ(S0) with standard deviation exponent
βrand = 1/2. This result can be interpreted in terms of the uncorrelated nature of the
randomized dataset (SI Section III). We consider that the population of each cell jincreases
by a random amount δjwith mean value ¯
δand variance h(δ−¯
δ)2i= ∆2, and that r≪1,
then n(i)
j(t1) = n(i)
j(t0) + δj. Therefore, the population of a cluster at time t1can be written
as
S1=S0+
Ni
X
j=1
δj.(4)
It can be shown that (SI Section III.):
hS2
1i=hS2
0i+
Ni
X
j
Ni
X
k
h(δj−¯
δ)(δk−¯
δ)i.(5)
Randomly shuffling population cells destroys the correlations, leading to h(δj−¯
δ)(δk−¯
δ)i=
∆2δjk (where δjk is the Kronecker delta function) which implies βrand = 1/2 [16] (see SI
Section III.).
The fact that βlies below the random exponent (βrand = 1/2) for all the analyzed data
7

suggests that the dynamics of the population cells display spatial correlations, which are
eliminated in the random surrogate data. The cells are not occupied randomly but spatial
correlations arise, since when the population in one cell increases, the probability of growth
in an adjacent cell also increases. That is, development attracts further development, an
idea that has been used to model the spatial distribution of urban patterns [17]. Indeed this
ideas are related to the study of the origin of power-laws in complex systems [18, 19].
When we analyze the populated cells, we indeed find that spatial correlations in the incre-
mental population of the cells, δj, are asymptotically of a scale-invariant form characterized
by a correlation exponent γ,
h(δj−¯
δ)(δk−¯
δ)i ∼ ∆2
|~xj−~xk|γ,(6)
where ~xjis the location of cell j. For GB we find γ= 0.93 ±0.08 (see Fig. 4D). In SI
Section III. we show that power-law correlations in the fluctuations at the cell level, Eq. (6),
lead to a standard deviation exponent β=γ/4. For γ= 2, the dimension of the substrate,
we recover βrand = 1/2 (larger values of γresult in the same βsince when γ > 2 correlations
become irrelevant). If γ= 0, the standard deviation of the populations growth rates has
no dependence on the population size (β= 0), as stated by Gibrat’s law, stating that the
standard deviation does not depend in the cluster size. In the case of GB, γ= 0.93 ±0.08
gives β= 0.23 ±0.02 approximately consistent with the measured value βGB = 0.27 ±0.04,
within the error bars. This observation suggests that the underlying demographic process
leading to the scaling in the standard deviation can be modeled as arising from the long-range
correlated growth of population cells.
III. DISCUSSION
Our results suggest the existence of scale-invariant growth mechanisms acting at different
geographical scales. Furthermore, Eq. (3) is similar to what is found for the growth of
firms and other macroeconomic indicators [16, 20]. Thus, our results support the existence
of an underlying link between the fluctuation dynamics of population growth and various
economic indicators, implying considerable unevenness in economic development in different
population sizes. City growth is driven by many processes of which population growth and
migration is only one. Our study captures only the growth of population but not economic
8

growth per se. Many cities grow economically while losing population and thus the processes
we imply are those that influence a changing population. Our assumption is that population
change is an indicator of city growth or decline and therefore we have based our studies on
population clustering techniques. Alternatively, the MSAs provides a set of rules that try
to capture the idea of city as a functional economic region.
The results we obtain show scale-invariant properties which we have modelled using long-
range spatial correlations between the population of cells. That is, strong development in
an area attracts more development in its neighborhood and much beyond. A key finding is
that small places exhibit larger fluctuations than large places. The implications for locating
activity in different places are that there is a greater probability of larger growth in small
places, but also a greater probability of larger decline. Opportunity must be weighed against
the risk of failure.
One may take these ideas to a higher level of abstraction to study cell-to-cell flows (mi-
gration, commuting, etc.) gridded at different levels. As a consequence one may define
population clusters, or MSAs, in terms of functional linkages between neighboring cells. In
addition one may relax some conditions imposed in the CCA. Here we consider a cell to be
part of a cluster only if its population is strictly greater than 0. In SI Section V we relax
this condition and study the robustness of the CCA when cells of a higher population than
0 (for instance, 5 and 20) are allowed into clusters and find that even though small clusters
present a slight deviation, the overall behavior of the growth rate and standard deviation is
conserved.
IV. MATERIALS AND METHODS
A. Information on the datasets
The datasets analyzed in this paper were obtained from the websites
http://census.ac.uk, http://www.esri.com/, and http://na.unep.net/datasets/datalist.php,
for GB, USA and Africa, respectively, and can be downloaded from
http://lev.ccny.cuny.edu/∼hmakse/cities/city data.zip.
The datasets consist of a list of populations at specific coordinates at two time steps t0
and t1. A graphical representation of the data can be seen in Fig. 1C for GB where each
9

point represents a data point directly extracted from the dataset.
To perform the CCA at different scales we coarse-grain the datasets. For this purpose,
we overlay a grid on the corresponding map (USA, GB, or Africa) with the desired cell size
(for example, 2km-by-2km or 4km-by-4km for the USA). Then, the population of each cell
is calculated as the sum of the populations of points (obtained from the original dataset)
that fall into this cell.
Table I shows information on the datasets and results on USA, GB and Africa for the
cell size used in the main text as well as some of the exponents obtained in our analysis.
TABLE I: Characteristics of datasets and summary of results
Data Number t0t1Average Cell Size Number of α β
of cells growth rate clusters
USA 1.86 mill 1990 2000 0.9% 2km-by-2km 30,210 0.28 ±0.08 0.20 ±0.06
GB 0.10 mill 1981 1991 0.3% 2.2km-by-2.2km 10,178 0.17 ±0.05 0.27 ±0.04
Africa 2,216 1960 1990 4% 77km-by-77km 3,988 0.21 ±0.05 0.19 ±0.04
B. Calculation of hr(S0)iand σ(S0)and methodology
The average growth rate, hr(S0)i= ln(S1/S0), and the standard deviation, σ(S0) =
phr(S0)2i − hr(S0)i2, are defined as follows. If we call P(r|S0) the conditional probability
distribution of finding a cluster with growth rate r(S0) with the condition of initial popula-
tion S0, then we can obtain r(S0) and σ(S0) through,
hr(S0)i=ZrP (r|S0)dr, (7)
and
hr(S0)2i=Zr2P(r|S0)dr. (8)
Once r(S0) and σ(S0) are calculated for each cluster, we perform a nonparametric re-
gression analysis [13, 14], a technique broadly used in the literature of population dynamics.
The idea is to provide an estimate for the relationship between the growth rate and S0and
between the standard deviation and S0. Following the methods explained in Ref. [14] we
10

apply the Nadaraya-Watson method to calculate an estimate for the growth rate, ˆr(S0),
with,
hˆr(S0)i=Pallclusters
i=0 Kh(S0−Si(t0))ri(S0)
Pallclusters
i=0 Kh(S0−Si(t0)) ,(9)
and an estimate for the standard deviation ˆσ(S0) with,
ˆσ(S0) = sPallclusters
i=0 Kh(S0−Si(t0))(ri(S0)− hˆr(S0)i)2
Pallclusters
i=0 Kh(S0−Si(t0)) ,(10)
where Si(t0) is the population of cluster iat time t0(as defined in the main text), ri(S0) is
the growth rate of cluster iand Kh(S0−Si(t0)) is a gaussian kernel of the form,
Kh(S0−Si(t0)) = e(lnS0−lnSi(t0))2
2h2, h = 0.5 (11)
Finally, we compute the 95% confidence bands (calculated from 500 random samples with
replacement) to estimate the amount of statistical error in our results [13]. The bootstrap-
ping technique was applied by sampling as many data points as the number of clusters and
performing the nonparametric regression on the sampled data. By performing 500 realiza-
tions of the bootstrapping algorithm and extracting the so called α/2 (αis not related to
the growth rate exponent) quantile we obtain the 95% confidence bands.
To obtain the exponents αand βof the power-law scalings for hr(S0)iand σ(S0), respec-
tively, we perform an OLS regression analysis [15]. More specifically, to obtain the exponent
βfrom Eq. (3), we first linearize the data by considering the logarithm of the independent
and dependent variables so that Eq. (3) becomes ln σ(S0)∼βln S0. Then, we apply a
linear Ordinary Least Square regression that leads to the exponent
β=NcPNc
i=1[ln Si(t0) ln σ(Si(t0))] −PNc
i=1 ln Si(t0)PNc
i=1 ln σ(Si(t0))
NcPNc
i=1(ln Si(t0))2−(PNc
i=1 ln Si(t0))2,(12)
where Ncis the number of clusters found using the CCA. Analogously, we obtain the expo-
nent αby linearizing h|r(S0)|i and calculating
α=NcPNc
i=1(ln Si(t0) ln h|r(Si(t0))|i − PNc
i=1 ln Si(t0)PNc
i=1 ln h|r(Si(t0))|i
NcPNc
i=1(ln Si(t0))2−(PNc
i=1 ln Si(t0))2.(13)
Next we compute the 95% confidence interval for the exponents αand β. For this we
follow the book of Montgomery and Peck [15]. The 95% confidence interval for βis given
by,
t0.025,Nc−2∗se, (14)
11

where tα′/2,Nc−2is the t-distribution with parameters α′/2 and Nc−2 and se is the standard
error of the exponent defined as
se =sSSE
(Nc−2)Sxx
,(15)
where SSEis the residual and Sxx is the variance of S0.
Finally, we express the value of the exponent in terms of the 95% confidence intervals as,
β±t0.025,Nc−2∗se. (16)
Acknowledgments
We thank L.H. Dobkins and J. Eeckhout for providing data on MSA and C. Briscoe
for help with the manuscript. This work is supported by the National Science Foundation
through grant NSF-HSD. J.S.A. thanks the Brazilian agencies CNPq, CAPES, FUNCAP
and FINEP for financial support.
12

[1] Gabaix X (1999) Zipf’s law for cities: an explanation. Quart. J. Econ. 114: 739-767.
[2] Gabaix X, Ioannides Y M (2003) The evolution of city size distributions, in Handbook of Urban
and Regional Economics, Vol.4, eds Henderson J V, Thisse J F (Elsevier Science, Amsterdam),
pp 2341-2378.
[3] Unwin D J (1996) GIS, spatial analysis and spatial statistics. Progress in Human Geography
20:540-551.
[4] King G, Rosen O, Tanner M A (2004) eds, Ecological Inference: New Methodological Strategies
(Cambridge University Press, New York).
[5] Eeckhout J (2004) Gibrat’s law for (all) cities. Amer. Econ. Rev. 94: 1429-1451.
[6] Dobkins L H, Ioannides Y M (2000) Spatial interactions among U.S. cities: 1900-1990. Reg.
Sci. Urban Econ. 31: 701-731.
[7] Ioannides Y M, Overman H G (2003) Zipf ’s law for cities: an empirical examination Reg. Sci.
Urban Econ. 33: 127-137.
[8] The 1981 and 1991 population census, Crown Copyright, ESRC purchase,
http://census.ac.uk/
[9] ESRI Inc (2000) ArcView 3.2 data sets: North America, Environmental Systems Research
Institute, Redlands, CA.
[10] UNESCO (1987) through GRID, http://na.unep.net/datasets/datalist.php
[11] Stauffer D (1984) Introduction to percolation theory (Taylor & Francis, London).
[12] Eaton J, Eckstein Z (1997) Cities and growth: theory and evidence from France and Japan.
Reg. Sci. Urban Econ. 27: 443-474.
[13] H¨ardle W (1990) Applied Nonparametric Regression (Cambridge University Press, Cam-
bridge).
[14] Silverman, B W (1986) Density Estimation for Statistics and Data Analysis (Chapman and
Hall, New York).
[15] Montogomery D C, Peck E A (1992) Introduction to linear regression analysis (John Wiley &
Sons, Inc.).
[16] Stanley M H R et al. (1996) Scaling behavior in the growth of companies. Nature 379: 804-806.
[17] Makse H A, Havlin S, Stanley H E (1995) Modelling urban growth patterns. Nature 377:
13

608-612.
[18] Barab´asi A-L, Albert R (1999) Emergence of scaling in random networks. Science, 286, 509-
512.
[19] Carlson J M, Doyle J (2002) Complexity and robustness. PNAS, 99:2538-2545.
[20] Rossi-Hansberg E, Wright M L J (2007) Establishment size dynamics in the aggregate econ-
omy. Amer. Econ. Rev. 97: 1639-1666.
14

FIG. 1: (A) Sketch illustrating the CCA applied to a sample of gridded population data. In the top
left panel, cells are colored in blue if they are populated (n(i)
j(t)>0), otherwise, if n(i)
j(t) = 0, they
are in white. In the top right panel we initialize the CCA by selecting a populated cell and burning
it (red cell). Then, we burn the populated neighbors of the red cell as shown in the lower left panel.
We keep growing the cluster by iteratively burning neighbors of the red cells until all neighboring
cells are unpopulated, as shown in the lower right panel. Next, we pick another unburned populated
cell and repeat the algorithm until all populated cells are assigned to a cluster. The population
Si(t) of cluster iat time tis then Si(t) = PNi
j=1 n(i)
j(t). (B) Cluster identified with the CCA in the
London area (red) overlaying a corresponding satellite image (extracted from maps.google.com).
The greenery corresponds to vegetation, and thus approximately indicates unoccupied areas. For
example, Richmond Park can be found as a vegetation area in the south-west. The areas in the
east along the Thames River correspond mainly to industrial districts and in the west the London
Heathrow Airport, also not populated. The yellow line in the center represents the administrative
boundary of the City of London, demonstrating the difference with the urban cluster found with the
CCA. The pink clusters surrounding the major red cluster are smaller conglomerates not connected
to London. The figure shows that an analysis based on the City of London captures only a partial
area of the real urban agglomeration. (C) Result of the CCA applied to all of GB showing the
large variability in the population distribution. The color bar (in logarithmic scale) indicates the
population of each urban cluster.
FIG. 2: Illustration of possible changes in cluster shapes. In each case we show how the growth
rate is computed. In the first case, there is no areal modification in the cluster between t0and t1.
In the second, the cluster expands. In the third the cluster reduces its area. In the fourth, one
cluster divides into two and therefore we consider the population at t1to be S1=S′
1+S′′
1. In the
fifth case two clusters merge to form one at t1. In this case we consider the population at t0to be
S0=S′
0+S′′
0.
15

FIG. 3: Results for the USA using a cell size of 2000m-by-2000m. (A) Mean annual growth rate for
population clusters in the USA versus initial population of the clusters. The straight dashed line
shows a power-law fit with αUSA = 0.28 ±0.08 as determined using OLS regression. (B) Standard
deviation of the growth rate for the USA. The straight dashed line corresponds to a power-law fit
using OLS regression with βUSA = 0.20 ±0.06.
FIG. 4: Results for Great Britain using a cell size of 2.2km-by-2.2km. (A) Mean annual growth
rate of population clusters in Great Britain versus the initial cluster population. The inset shows
a double logarithmic plot of the growth rate in the intermediate range of populations, 3000 <
S0<3×105. A power-law fit using OLS leads to an exponent αGB = 0.17 ±0.05 for this range.
(B) Double logarithmic plot of the standard deviation of the annual growth rates of population
clusters in Great Britain versus the initial cluster population. The straight line corresponds to a
power-law fit using OLS with an exponent βGB = 0.27 ±0.04, according to Eq. (3). (C) Scaling
of the standard deviation in cluster population obtained from the randomized surrogate dataset
of GB by randomly swapping the cells. The data shows an exponent βrand = 1/2 in the tail. The
deviations for small S0are discussed in the SI Section IV. where we test these results by generating
random populations. (D) Long-range spatial correlations in the population growth of cells for GB
according to Eq. (6). The straight line corresponds to an exponent γ= 0.93 ±0.08.
FIG. 5: Results for Africa using a cell size of 77km-by-77km. (A) Mean growth rate of clusters in
Africa versus the initial size of population S0. The straight dashed line shows a power-law fit with
exponent αAf = 0.21 ±0.05, obtained using OLS regression. (B) Standard deviation of the growth
rate in Africa. The straight line corresponds to power-law fit using OLS providing the exponent
βAf = 0.19 ±0.04.
16

A
Fig. 1:
B
Fig. 1
17

C
Fig. 1
18

t0t1
No Change
Expansion
Reduction
Division
Merge
S0
S0
‘
‘
S1
‘
S0S1
S0S1
S1
‘
‘
S0
‘
S1
r(S0) = ln S1
S0
r(S0) = ln S1
S0
r(S0) = ln S1
S0
r(S0) = ln S1+S1
S0
r(S0) = ln S1
S0+S0
S0=S0+S0
S1=S1+S1
S0S1
Fig. 2:
19

A B
FIG. 3:
20

A B
C D
FIG. 4:
21

A B
FIG. 5:
22

SUPPORTING INFORMATION
Laws of Population Growth
Hern´an D. Rozenfeld, Diego Rybski, Jos´e S. Andrade Jr.,
Michael Batty, H. Eugene Stanley, and Hern´an A. Makse
As supplementary materials we provide the following: In Section V we present tables
with details on our results using the CCA and results presented in previous papers to allow
for comparison between the different approaches. In Section VI we study the stability of the
scaling found in the text under a change of scale in the cell size. In Section VII we detail the
calculations to relate spatial correlations between the population growth and σ(S0) namely
the relation β=γ/4. In Section VIII we describe the random surrogate dataset used to
further test our results. In Section IX we further test the robustness of the CCA by proposing
a small variation in the algorithm.
V. CLUSTERS AT DIFFERENT SCALES AND COMPARISON WITH
METROPOLITAN STATISTICAL AREAS
In this section, Tables S1 and S2 allow for a detailed comparison of urban clusters obtained
with the CCA applied to the USA in 1990, and the populations of MSA from US Census
Bureau used in previous studies of population growth [5, 6, 7].
We can see that the MSA presented by Eeckhout (2004) typically correspond to our
clusters using cell sizes of 4km and 8km. For example, for the New York City region
Eeckhout’s data are well approximated by a cell size of 4km, but Los Angeles is better
approximated when using a cell size of 8km. On the other hand Dobkins-Ioannides (2000)
data are better described by cell sizes of 2km or 4km. For instance, Chicago is well described
by a cell size of 4km and Los Angeles is better described by a cell size of 2km.
An interesting remark is that the population of Los Angeles when using cell sizes of 2km,
4km and 8km does not vary as much as that for New York. This could be caused by the
fact that major cities in the northeast of USA are closer to each other than large cities in
the southwest, which may be attributed to land or geographical constraints.
23

It is important relate the results of Table S2 with an ecological fallacy. As the cell size
is increased, the population of a cluster also increases, as expected, because the cluster now
covers a larger area. This is not a direct manifestation of an ecological fallacy which, would
appear if the statistical results (growth rate vs. S or standard deviation vs. S) gave different
results as the cell size increases. In Fig. 1 and Fig. 2 in the SI Section VI, we observe that
the growth rate and standard deviation for the USA and GB follow the same form, except
for the case of the growth rate in the USA in which different cell sizes show deviations from
each other. The later may be an indicative of an ecological fallacy. In this case, it is not
obvious what cell size is the “correct” one. We consider this point (the possibility to choose
the cell size) to be a feature of the CCA, since one may appropriately pick the cell size
according to the specific problem one is studying.
Table S1: Top 10 largest MSA of the USA in 1990 from previous analysis of
population growth
Dobkins - Ioannides Eeckhout
MSA Population MSA Population
1 NYC NY206 9,372,000 NYC-North NJ-Long Is., NY-NJ-CT-PA 19,549,649
2 Los Angeles CA172 8,863,000 Los Angeles-Riverside-Orange County, CA 14,531,529
3 Chicago IL59 7,333,000 Chicago-Gary-Kenosha, IL-IN-WI 8,239,820
4 Philadelphia PA228 4,857,000 Washington-Baltimore, DC-MD-VA-WV 6,727,050
5 Detroit MI80 4,382,000 San Francisco-Oakland-San Jose, CA 6,253,311
6 Washington DC312 3,924,000 Philadelphia-Wilmington-Atlantic City 5,892,937
PA-NJ-DE-MD
7 San Francisco CA266 3,687,000 Boston-Worcester-Lawrence, MA-NH-ME-CT 5,455,403
8 Houston TX129 3,494,000 Detroit-Ann Arbor-Flint, MI 5,187,171
9 Atlanta GA19 2,834,000 Dallas-Fort Worth, TX 4,037,282
10 Boston MA39 2,800,000 Houston-Galveston-Brazoria, TX 3,731,131
24

Table S2: Top 10 largest clusters of the USA in 1990 from our analysis for
different cell sizes. The city names are the major cities that belong to the clusters and
were picked to show the areal extension of the cluster.
Cell = 1km Cell = 2km Cell = 4km Cell = 8km
Cluster Population Cluster Population Cluster Population Cluster Population
1 NYC 7,012,989 NYC-Long Is. 12,511,237 NYC-Long Is. 17,064,816 NYC-Long Is. 41,817,858
Newark N. NJ-Newark North NJ
Jersey City Jersey City Philadelphia
D.C.-Boston
2 Chicago 2,312,783 Los Angeles 9,582,507 Los Angeles 10,878,034 Los Angeles 13,304,233
Long Beach Long Beach San Clemente
Pomona Riverside
3 Los Angeles 1,411,791 Chicago 4,836,529 Chicago 7,230,404 Chicago 9,288,345
Rockford Gary Gary
Rockford Rockford
Milwaukee
4 Philadelphia 1,282,834 Philadelphia 3,151,704 Washington 5,316,890 San Francisco 5,736,479
Wilmington Baltimore Santa Cruz
Springfield Brentwood
5 Boston 759,024 Detroit 2,906,453 Philadelphia 4,935,734 Detroit 4,442,723
Trenton Ann Arbor
Wilmington Monroe
Sarnia
6 Newark 581,048 San Francisco 2,601,639 San Francisco 4,766,960 Miami 4,000,432
San Jose San Jose Port St. Lucie
Concord
7 San Francisco 507,300 Washington 2,059,421 Detroit 3,722,778 Dallas 3,536,186
Alexandria Waterford Fort Worth
Bethesda Canton
8 Washington 504,068 Phoenix 1,556,077 Miami 3,719,773 Houston 3,425,647
W. Palm Beach
9 Jersey City 438,591 Boston 1,498,208 Dallas 3,134,233 Cleveland 3,233,341
Lowell Fort Worth Canton
Quincy
10 Baltimore 437,413 Miami 1,465,490 Boston 3,064,925 Pittsburgh 3,214,661
Brockton Youngstown
Nashua Morgantown
25

A B
FIG. 6: Sensitivity of the results under coarse-graining of the data for GB. (A) Average growth
rate and (B) standard deviation for GB using the clustering algorithm for different cell size. The
dashed line represents the OLS regression estimate for the exponents (A) αGB = 0.17 and (B)
βGB = 0.27 obtained in the main text. For clarity we do not show the confidence bands.
VI. SCALING UNDER COARSE-GRAINING
In this section we test the sensitivity of our results to a coarse-graining of the data. We
analyze the average growth rate hr(S0)iand the standard deviation σ(S0) for GB and the
USA by coarse-graining the data sets at different levels.
In Fig. 6A we observe that although the results are not identical for all coarse-grainings,
they are statistically similar, showing a slight decay in the growth rate. Moreover, we see
that cities of size S0≈103and S0≈106still exhibit a tendency to have negative growth
rates for all levels of coarse-graining, as explained in the main text. In the case of the USA
(Fig. 7A) there is a crossover to a flat behavior at a cell size of 8000m, although at this scale
all the northeast USA becomes a large cluster of 41 million inhabitants. On the other hand,
Figs. 6B, 7B show that the scaling of Eq. (3) in the main text, σ(S0)∼S−β
0, still holds when
using the coarse-grained datasets on both GB and the USA.
VII. CORRELATIONS
In this section we elaborate on the calculations leading to the relation between Gibrat’s
law and the spatial correlations in the cell population. We first show that when the pop-
26

A B
FIG. 7: Study of results under coarse-graining of the data for the USA. (A) Average growth rate
and (B) standard deviation for the USA using the clustering algorithm for different cell size. The
dashed line represents the OLS regression estimate for the exponents (A) αUSA = 0.28 and (B)
βUSA = 0.20 obtained in the main text. For clarity we do not show the confidence bands.
ulation cells are randomly shuffled (destroying any spatial correlations between the growth
rates of the cells), the standard deviation of the growth rate becomes σ(S0)∼S−βrand
0, where
βrand = 1/2 [16]. Then, we show that long-range spatial correlations in the population of
the cells leads to the relation β=γ/4 as stated at the end of Section II in the main text.
Assuming that the population growth rate is small (r≪1), we can write R=er≈1 + r.
Replacing R= 1 + rin Eq. (1) in the main text we obtain
S1=S0+S0r. (17)
We define the standard deviation of the populations S1as σ1, which is a function of S0:
σ1(S0) = qhS2
1i − hS1i2.(18)
This quantity is easier to relate to the spatial correlations of the cells than the standard
deviation σ(S0) of the growth rates r. Then, since hS1i=S0+S0hriand hS2
1i=S2
0+
2S2
0hri+S2
0hr2i, we obtain,
σ1(S0)∼S0σ(S0),(19)
where σ(S0) = phr2i − hri2as defined in the main text. Therefore, using Eq. (3) in the
27

main text,
σ1(S0)∼S1−β
0.(20)
As stated in the main text, the total population of a cluster at time t0is the sum of the
populations of each cell, S0=PNi
j=1 n(i)
j, where Niis the number of cells in cluster i. The
population of a cluster at time t1can be written as
S1=S0+
Ni
X
j=1
δj,(21)
where δjis the increment in the population of cell jfrom time t0to t1(notice that δjcan
be negative). Therefore, the standard deviation σ1(S0) is
σ1(S0)2=
Ni
X
j,k
hδjδki − h
Ni
X
j
δji2=
Ni
X
j,k
h(δj−¯
δ)(δk−¯
δ)i.(22)
After the process of randomization explained in Section II main text, the correlations
between the increment of population in each cell are destroyed. Thus,
h(δj−¯
δ)(δk−¯
δ)i= ∆2δjk ,(23)
where ∆2=¯
δ2−¯
δ2. Replacing in Eq. (22) and since hni= (1/Ni)PNi
jnj=S0/Ni, we
obtain
σ1(S0)2=Ni∆2∼S0.(24)
Comparing with Eq. (20) we obtain βrand = 1/2 for this uncorrelated case.
Let us assume that the correlation of the population increments δj, decays as a power-law
of the distance between cells indicating long-range scale-free correlations. Thus, asymptoti-
cally
h(δj−¯
δ)(δk−¯
δ)i ∼ ∆2
|~xj−~xk|γ,(25)
where ~xjdenotes the position of the cell jand γis the correlation exponent (for |~xj−~xk| → 0,
the correlations h(δj−¯
δ)(δk−¯
δ)itend to a constant). For large clusters, we can approximate
the double sum in Eq. (22) by an integral. Then, assuming that the shape of the clusters
can be approximated by disks of radius rc, for γ < 2 we obtain
(σ1(S0))2=
Ni
X
j,k
∆2
|~xj−~xk|γ→∆2Ni
a2Zrcrdrdθ
rγ≈∆2
(2 −γ)
Ni
a2r−γ+2
c,(26)
28

where a2is the area of each cell and rcthe radius of the cluster. Since rc∼Nia2, we finally
obtain,
σ1(S0)2∼N2−γ
2
i.(27)
Using S0=Nihniand Eq. (20) we arrive at,
β=γ
4.(28)
Equation (28) shows that Gibrat’s Law is recovered when the correlation of the population
increments is a constant, independent from the positions of the cells; that is when all the
populations cells are increased equally. In other words, if γ= 0, the standard deviation of
the populations growth rates has no dependence on the population size (β= 0), as stated by
Gibrat’s law. The random case is obtained for γ=d, where d= 2 is the dimensionality of the
substrate. In this case d= 2 and βrand = 1/2. For γ > 2, the correlations become irrelevant
and we still find the uncorrelated case βrand = 1/2. For intermediate values 0 < γ < 2 we
obtain 0 < β =γ/4<1/2.
VIII. RANDOM SURROGATE DATASET
In this section we elaborate on the randomization procedure used to understand the role
of correlations in population growth.
Figure 4C in the main text shows the standard deviation σ(S0) when the population
of each cluster is randomized, breaking any spatial correlation in population growth. For
clusters with a large population, σ(S0) follows a power-law with exponent βrand = 1/2,
and for small S0,σ(S0) presents deviations from the power-law function as seen in Fig. 4C
with smaller standard deviation than the prediction of the random case. This deviation is
caused by the fact that the population of a cluster is bound to be positive: a cluster with a
small population S0cannot decrease its population by a large number, since it would lead
to negative values of S1. This produces an upper bound in fluctuations of the growth rate
for small S0and results in smaller values of σ(S0) than expected (below the scaling with
exponent βrand = 1/2).
To support this argument, we carry out simulations using the clusters of GB, where the
population nj(t0) of each cell jis replaced with random numbers following an exponential
distribution with probability P(nj)∼e−nj/n0. The decay-constant, n0= 150, is extracted
29

from the data of GB to mimic the original distribution. This is done through a direct measure
of P(nj) from the GB dataset and fitting the data using OLS regression analysis. We obtain
the population nj(t1) = nj(t0) + δjof cell jat time t1by picking random numbers for the
population increments δjfollowing a uniform distribution in the range −q∗150 < δi< q∗150.
Here qdetermines the variance of the increments. Since the population cannot be negative
we impose the additional condition nj(t1)≥0. Figure 8 shows the results of the standard
deviation σ(S0) for four different q-values for this uncorrelated model. We find that the tail
of σ(S0) reproduces the uncorrelated exponent βrand = 1/2. For small S0we find that the
standard deviation levels off to an approximately constant value as in the surrogate data of
Fig 4C. The crossover from an approximately constant σ(S0) to a power-law moves to smaller
values of the population S0as the standard deviation in the δjis smaller (smaller value of q).
Such behavior can be understood since the condition n(i)
j(t1)≥0 imposes a lower “wall” in
the random walk specified by n(i)
j(t1) = n(i)
j(t0) + δj. As the initial population gets smaller,
the walker “feels” the presence of the wall and the fluctuations decrease accordingly, thus
explaining the deviations from the power-law with exponent βrand = 1/2 for small population
values. Therefore, as the value of qdecreases, the small population plateau disappears as
observed in Fig. 8.
IX. A VARIATION OF THE CCA
In this section we study a variation of the CCA. In the main text we stop growing a cluster
when the population of all boundary cells have unpopulated, that is, have population exactly
0. In other words, clusters are composed by cell with population strictly greater than 0. It
is important to analyze whether this stopping criteria can be relaxed to including cell which
have a population larger that a given threshold. In Fig. 9A and Fig. 9B we show the results
for the population growth rate and standard deviation, respectively, in GB when the cell
size is 2.2km-by-2.2km (as in the main text) but including cells with a population strictly
larger than 5 and 20.
Although for small population clusters we observe a slight variation in the growth rate
and in the standard deviation, the results show that the thresholds do not influence the
global statistics when compared to the plots in the main text.
30

FIG. 8: Standard deviation σ(S0) for the random data set as explained in the SI Section VIII.
The results for σ(S0) are rescaled to collapse the power-law tails with exponent βrand = 1/2 and to
emphasize the deviations from this function for small values of S0. The larger the parameter q, the
larger the deviations from the power-law at lower S0. In other words, the crossover to power-law
tail appears at larger S0as qincreases.
AB
FIG. 9: Sensitivity of the results under a change in the stopping criteria in the CCA (A) Average
growth rate for GB with a population threshold of 5 (green line) and 20 (black dashed line) and
(B) standard deviation for GB with a population threshold of 5 (green line) and 20 (black dashed
line). For clarity we do not show the confidence bands.
31