1 
 
The Distance-Decay Function of Geographical Gravity Model: 
Power Law or Exponential Law? 
 
Yanguang Chen 
(Department of Geography, College of Urban and Environmental Sciences, Peking University, Beijing 
100871, P.R. China. E-mail: chenyg@pku.edu.cn) 
 
Abstract: The distance-decay function of the geographical gravity model is originally an inverse 
power law, which suggests a scaling process in spatial interaction. However, the distance exponent 
of the model cannot be explained with the ideas from Euclidean geometry. This results in what is 
called dimension dilemma. In particular, the gravity model based on power law could not be 
derived from general principles by traditional ways. Consequently, a negative exponential function 
substituted for the inverse power function to serve for a distance-decay function for the gravity 
model. However, the exponential-based gravity model goes against the first law of geography. 
This paper is devoted to solve these kinds of problems by mathematical reasoning and empirical 
analysis. First, it can be proved that the distance exponent of the gravity model is essentially a 
fractal dimension. Thus the dimensional dilemma of the power-based gravity model can be 
resolved using the concepts from fractal geometry. Second, the exponential function indicates 
locality and localization, which violates the basic principle of spatial interaction. The power 
function implies action at a distance, which is the necessary condition of geographical gravitation. 
Third, the gravity model based on power law decay can be derived from the entropy- maximizing 
principle by introducing a proper postulate. The observational data of China’s cities and regions 
are employed to verify the theoretical inferences, and the results support power-law distance decay. 
A conclusion can be reached that the preferred form of geographical gravity model is its original 
form, which is based on an inverse power law rather than a negative exponential law.  
 
Key words: gravity model; distance-decay function; allometry; fractals; fractal dimension; spatial 
interaction; dimension dilemma; entropy-maximization 

2 
 
Many of our theories in physical and human geography are being reinterpreted using ideas from 
fractals and tomorrow they will become as much a part of our education and experience as maps 
and statistics are today. 
Michael Batty (1992) 
1 Introduction 
The gravity model is one of basic models of spatial analysis in geography and social physics. It 
provides an empirically effective approach to modeling spatial interaction. The model is originally 
proposed to describe population migration between two regions (Carey, 1858; Grigg, 1977; 
Ravenstein, 1885). Afterward, it is employed to measure the force of attraction between any two 
geographical objects such as cities, firms, and retail stores. Today, the model can be found in many 
subjects like economics and sociology. A lot of variants of the model came out, and different 
forms of gravity models have different spheres of application (Erlander, 1980; Haynes and 
Fotheringham, 1984; Sen and Smith, 1995). All the models remain essentially the same despite 
apparent changes. The key rests with the law of distance decay of spatial interaction. Despite its 
practical effect, many basal problems are still pending and remain to be further explored. As is 
known, the geographical gravity model is initially found by analogy with Newton’s law of 
universal gravitation (Wilson, 2000), thus its distance-decay function is an inverse power function. 
Nowadays, the inverse power law plays an important role in the studies on complexity. However, 
for traditional geography, the power law is an obstruction in the way of theoretical construction. 
A power function indicates a relation of proportionality between two measures such as length, 
area, and volume. The geometric measure relations result in a principle of dimension consistency. 
Two quantities are proportional to one another if and only if they bear the same dimension. The 
Euclidean dimension must be an integer (0, 1, 2, and 3). Unfortunately, many power laws and 
power-based models such as the law of allmetric growth and the original gravity model defy this 
dimension principle. By dimensional analysis, the distance exponent of the gravity model should 
be an integer or a ratio of one integer to another integer. However, large numbers of empirical 
results of the distance exponent failed to support the values predicted by the axiom of dimensional 
homogeneity. The observed values of the exponent are always neither integers nor ratios of two 

3 
 
integers, but random values coming between 0 and 3. This cannot be interpreted by the ideas from 
the traditional mathematics based on Euclidean geometry. Therefore the gravity model once fell in 
a dimensional dilemma. To solve the dimension problem, theoretical geographers replaced the 
power function with an exponential function to serve as a distance-decay function (Haggett et al, 
1977; Haynes, 1975). This model came from the spatial interaction model derived from 
entropy-maximization principle by Wilson (1970). In terms of the viewpoint of geographers in 
those years, Wilson’s work provided a firmer theoretical basis for many of the spatial interaction 
models based on loose physical analogy and “raises the gravity model phoenix-like from the ashes” 
(Gould, 1972, p696; Haggett et al, 1977, page 47). 
However, the exponential-based gravity model caused a new problem. An inverse power 
function suggests an action at a distance, it is proper to reflect the nature of spatial interaction; 
while a negative exponential function suggests locality and localization, it cannot reflect the 
interaction between two distant places (Chen, 2008; Chen, 2012). In a sense, the function of 
exponential decay violates the first law of geography, which was presented by Tobler (1970; 2004). 
On the other hand, the dimensional problem of the distance exponent can be solved by the ideas 
from fractal theory (Chen, 2009), which was originated by Mandelbrot (1983). Today, many 
models based on power laws can be extricated from the dilemma of dimension (Batty and Longley, 
1994). The novel distance-decay analyses of recent years relate to power function rather than 
exponential function (Batty, 2008; Chen, 2009; Rybski et al, 2013). Partial sporadic conclusions 
has been published, by so far, there have been no systematic discussion on distance decay effect of 
the geographical gravity. This paper is devoted to clarifying the mathematical form of the 
distance-decay function and reinterpreting the distance exponent with fractal dimension concepts. 
In Section 2, a new theoretical framework is presented to explain the power-based gravity model. 
This theoretical framework is constructed using the ideas from fractals and entropy maximization. 
In Section 3, an empirical analysis of Chinese regions and cities is made to support the theoretical 
inferences on the fractal gravity model. In Section 4, several related questions are clarified or 
discussed. Finally, the paper is concluded by summarizing the main points of this study. 

4 
 
2 Theoretical models 
2.1 The fractality of the gravity model 
The distance exponent of the geographical gravity model can be interpreted with the concepts of 
scaling and fractal dimension. Suppose that there exist two places, which are numbered by i and j. 
According to first law of geography (Tobler, 1973), the two places are attracted to each other. The 
interaction between the two places can be measured with the gravity model 
b
ij
j
i
ij
r
P
P
G
I =
,                                  (1) 
where Iij denotes the gravity between places i and j, which can be represented with the quantity of 
the flow from one place to the other, Pi and Pj are the “mass”, which can be reflected by the 
population size of places i and j, rij is the distance between i and j, G refers to a proportionality 
coefficient, and b, to the distance exponent. Equation (1) is the original form of the gravity model. 
Suppose there is a point numbered x between places i and j. The gravity between i and x and that 
between j and x can be expressed as 
b
i
x
i
ix
r
P
P
G
I
=
, 
b
j
x
j
jx
r
P
P
G
I
=
.                            (2) 
where Iix denotes the gravity between i and x, Ijx is the gravity between j and x, ri refers to the 
distance from x to i, and rj to the distance from x to j. From equations (2) it follows 
jx
ix
b
j
j
b
i
i
I
I
r
P
r
P
=
/
/
.                                 (3) 
If Iix=Ijx as given, then 
b
j
i
j
i
r
r
P
P
)
(
=
,                                  (4) 
which is isomorphic to the law of retail gravitation derived by Reilly (1931). Reilly’s model was 
revised by Converse (1949), who derived a well-known breaking-point formula from equation (3). 
A breaking point can be defined as a balanced point between two places. The gravity from the 
breaking point to one place is equal to that from the point to the other place. 
A classical problem is how to interpret the distance exponent, b. Clearly, equation (3) indicates 
a spatial scaling relation. The distances r is a length, which denotes a lineal measure with a 

5 
 
dimension d=1. Let the dimension of the size measure P equal D. According to the geometric 
measure relation, we have 
d
j
i
D
j
i
r
r
P
P
/
1
/
1
)
(
)
(
=
,                                (5) 
where d is the Euclidean dimension of the distance. If P is a 2-dimensional measure, i.e., D=2, 
then we will have b=D/d=2; if the dimension of P is D=3, then b=D/d=3. However, in empirical 
work, the calculated value of b is neither 2 nor 3, but a random fraction comes between 0 and 3. 
This phenomenon cannot be interpreted using the concept from Euclidean geometry and gave rise 
to the dimension dilemma (Chen, 2009; Haggett et al, 1977; Haynes, 1975). In order to get over 
the conundrum, an exponential-based gravity model is proposed as follows 
0
/ r
r
j
i
ij
ij
e
P
GP
I
−
=
,                                (6) 
where r0 is a scale parameter of spatial interaction. The advantages of the exponential gravity 
model are as below: first, it is independent of dimension; second, its underlying rationale is clear 
because it is derivable from the principle of entropy maximization (Wilson, 1970).  
A gravity model is actually based on the action at a distance. However, the exponential-based 
model does not accord with the principle of spatial interaction. The negative exponential function 
bears a characteristic scale, r0, but it has no long memory. The characteristic scale is associated 
with the mean of distances. It can be demonstrated that an exponential decay suggests a locality 
and a process of localization rather than action at a distance (Chen, 2008). Thus a new problem 
arose from the substitution of the exponential distance-decay function for the power-law 
distance-decay function. Today, the dimension problem of power-based model can be easily 
solved by using the ideas from fractals. If the dimension of the size measure P is regarded as 
fractional, the observed values of b and be explained. For a fractal object, the dimension D can be 
arbitrary number varying between 0 and 3. Thus the distance exponent b=D/d=D will range from 0 
to 3, that is, 0<b<3. Another problem is the underlying rationale of the power-based gravity model. 
In other word, whether or not it can be derived from a general principle. In fact, the gravity model 
based on the power law can also be inferred with the method of entropy maximization. 
2.2 Derivation of the power-law based spatial interaction model 
The gravity model can be derived from the spatial interaction model, which can be derived from 

6 
 
the principle of entropy maximization. Wilson (1970) derived an exponential-based spatial 
interaction model using the entropy-maximizing method. In fact, the spatial interaction model 
based on power law can also be derived with the ideas from entropy maximization if we revise 
one of the postulates given by Wilson (1970), who assumed that the transport cost is linearly 
proportional to the corresponding distance. Suppose there is a geographical region which can be 
divided into n zones (sub-regions). The center of gravity of a zone can be represented by its 
central city. The maximum number of spatial flows between these zones is n×n. Let the total 
quantity of the flow be T, and the flow from zone i to zone j be Tij. Based on the assumption of 
regional partition, the state number of spatial flows is as follows 
∏∏
=
=
⎟⎟
⎠
⎞
⎜⎜
⎝
⎛
=
i
j
ij
nn
nn
ij
T
T
T
T
T
T
T
T
T
T
T
T
T
W
!
!
!
!
!
!
!
)
(
21
12
11
21
12
11
L
L
L
L
,      (7) 
where i、j=1,2,…,n. Based on the distribution of flows, the state entropy can be defined as 
∑∑
∏∏
−
=
−
=
=
i
j
ij
i
j
ij
ij
T
T
T
T
T
W
H
!
ln
!
ln
!
ln
!
ln
)
(
ln
.                (8) 
Suppose that the total outflow quantity from the ith origin, zone i, is Oi, and the total inflow 
quantity to the jth destination, zone j, is Dj, and the flow quantity from zone i to zone j is Tij. We 
have two constraint equations: ∑Tij=Oi, ∑Tij=Dj, from which it follows 
T
D
O
T
j
j
i
i
i
j
ij
=
=
=
∑
∑
∑∑
.                          (9) 
Then if the number n is a finite value, the maximizing process of the state entropy will result in a 
uniform distribution of spatial flows, namely, Tij≡const, where const denotes a constant. 
However, if the spatial differentiation of a geographical region is taken into consideration, the 
result will be different. In theory, if n→∞, then the distance between the centroid of one zone and 
that of another zone varies from 0 to infinity. The postulate of infinite number of zones is just a 
mathematical skill, by which a discrete variable can be converted into a continuous variable (Casti, 
1996). In virtue of the continuous variables, a physical/geographical problem can be solved by 
means of calculus. For simplicity, the variable of distance will be replaced with transportation cost. 
Different distances result in different costs. The long a distance is, the more the transport cost will 
be. Suppose the unit cost of the flow from zone i to zone j is cij, and the total cost of all the flow in 
this region is C. Equation (9) can be substituted and we have the third constraint equation 

7 
 
∑∑cijTij=C. Assuming the evolutive aim of the regional system is entropy maximization of the 
flow distribution, we can construct a nonlinear programming model, which was propounded by 
Wilson (1970). The objective function of the optimization problem is 
∑∑
−
=
i
j
ij
T
T
H
!
ln
!
ln
max    
,                           (10) 
which bears three equality constraints of limited inflow, outflow, and total cost of transportation. 
The constraint conditions can be expressed as 
⎪
⎪
⎪
⎩
⎪⎪
⎪
⎨
⎧
=
=
=
∑∑
∑
∑
i
j
ij
ij
j
i
ij
i
j
ij
C
T
c
D
T
O
T
      
s.t.
     
          
.                             (11) 
Finding the solution to equation (10) subject to equations (11) yields the relation between the flow 
quantity and transport cost. Thus spatial interaction models can be elegantly derived from the 
programming equations. An approach to solving the optimization problem is to utilize the 
Lagrange multiplier. Based on equations (10) and (11), a Lagrange function can be constructed as 
)
(
)
(
)
(
!
ln
!
ln
)
(
∑∑
∑
∑
∑
∑
∑∑
−
+
−
+
−
+
−
=
i
j
ij
ij
j
i
ij
j
j
i
j
ij
i
i
i
j
ij
ij
T
c
T
T
D
y
T
O
x
T
T
T
L
β
.(12) 
where xi, yj, and β refers to the Lagrange multipliers. The well-known Stirling’s formula, 
T!=(2π)1/2TT+1/2eT, can be employed to solve equation (12). If Tij is large enough, an approximate 
relation can be derived from Stirling’s formula, i.e., dln(Tij!)/dTij =ln(Tij). Taking the partial 
derivative of equation (12) with respect to Tij yields 
ij
j
i
ij
ij
ij
c
y
x
T
T
T
L
β
−
−
−
−
=
∂
∂
ln
)
(
.                        (13) 
The process of mathematical transformation is based on Stirling’s approximate relation given 
above. According to the related knowledge of calculus, the extremum condition is ∂L(Tij)/ ∂Tij =0. 
Thus equation (13) can be converted into an exponential function 
ij
j
i
c
y
x
ij
e
e
e
T
β
−
−
−
=
,                               (14) 
This equation leads to the family of Wilson’s spatial interaction models, which are based on the 
law of exponential distance decay (Wilson, 1970; Wilson, 2000). 
The deficiency of Wilson’s models mainly is reflected in two aspects. First, the negative 
exponential decay is not consistent with action at a distance (Chen, 2008). Exponential distance 

8 
 
decay defies the first law of geography. Second, geographical facts fail to support the linear 
relationship between traffic costs and distances. As mentioned above, one of Wilson’s assumptions 
is that the transport costs vary directly as the distances. However, this assumption does not accord 
with reality. As Haggett (2001, page 400) pointed out: “A more realistic curve for distance costs is 
convex and non-linear, indicating that transport costs increase, but at a decreasing rate, with 
distance.” The most probable curve of distance costs should confirm to a logarithmic function as 
ij
ij
r
c
c
c
ln
0
1 +
=
.                               (15) 
where c0 and c1 are two constants, and rij is the distance between the centers of gravity of zones i 
and j. Apparently, the growth rate of the transport cost is in inverse proportion to distance, that is, 
dcij/drij=c0/rij. This suggests that the transport cost increases over distance, but the rate of growth 
is decreasing. This accords with what is actually happening in the real world. Substituting 
equation (15) into equation (14) yields 
σ
σ
β
−
−
−
−
−
−
=
=
ij
y
x
c
ij
y
x
ij
r
e
e
r
r
r
e
e
T
i
i
j
i
0
0
0
)
(
,                     (16) 
where σ=βc0 refers to distance exponent, and r0=exp(-c1/c0) to the characteristic scale of distance, 
which relates to a mean distance. Without loss of generality, let 
i
i
x
O
A
e
d
i =
−
2
/
0
σ
, 
j
j
y
D
B
e
d
i =
−
2
/
0
σ
.                      (17) 
Thus a power-based spatial interaction model is gracefully derived from equations (10) and (11) as 
below 
σ
−
=
ij
j
i
j
i
ij
d
D
O
B
A
T
,                              (18) 
which was once empirically applied to transport analysis (Taylor, 1977). Differing from Wilson’s 
original model, the power-based model indicates the action at a distance rather than locality. 
Substituting equation (18) into the constraint equations ∑Tij=Oi and ∑Tij=Dj, which is inside 
equations (11), yields 
i
ij
j
j
i
j
i
O
d
D
O
B
A
=
−
∑
σ
, 
j
ij
i
j
i
j
i
D
d
D
O
B
A
=
−
∑
σ
.                 (19) 
Thus we obtain the following parameters 
)
/(
1
σ
−
∑
=
ij
j
j
j
i
d
D
B
A
, 
)
/(
1
σ
−
∑
=
ij
i
i
i
j
d
O
A
B
.                  (20) 
which are the scaling factors of spatial interaction defined by Wilson (1970). All the above 

9 
 
derivation is same with that presented by Wilson (1970) but a key postulate and the final result. 
According to the general knowledge of geography, the transport cost is not in linear proportion to 
distance. The new postulate implies that the growth rate of transport cost with distance is inversely 
proportional to distance. By improving a simple postulate, we have another spatial interaction 
model, which can be employed to explain and predict reality in a better way. 
2.3 Derivation of the power-law based gravity model 
A set of power-law based gravity models can be derived from the revised spatial interaction 
model through allometric relations. Allometric scaling is a ubiquitous phenomenon in both natural 
and human systems (Chen, 2014a). Bertallanfy (1968) once derived an aprior allometry from the 
principle of general system theory. In fact, allometric relation is a geometric measure relation 
(Chen, 2010a). Every system is organized according to certain proportions, and each proportion is 
associated with a geometric measure relation. Thus, the most probable relation between two 
geographical measures is the allometric relation (Chen and Jiang, 2009). Suppose that the relations 
between the inflow/outflow and population sizes follow the law of allometric scaling, that is 
u
i
i
P
O
η
=
,
v
j
j
P
D
μ
=
.                              (21) 
where η and μ refers to proportionality coefficients, and u and v to scaling exponents. Substituting 
equations (21) into equation (18) yields a gravity model such as 
σ
−
=
ij
v
j
u
i
ij
r
P
KP
T
,                                (22) 
where the coefficient K=ημAiBj. This the simple form of the gravity model. The attraction force is 
measured with spatial flows. Equation (22) has been empirically verified by observational data 
(Machay, 1959). A new finding is that there exists a symmetric form of equation (22) in the follow 
form 
σ
−
=
ij
u
j
v
i
ji
r
P
KP
T
,                               (23) 
where has been overlooked for a long time because that geographers takes it for granted that the 
output Tij equals input Tji. Actually, if and only if u=v=1, there is a possible symmetric distribution 
of flows and Tij=Tji (Batty and Karmeshu, 1983). Generally speaking, Tij≠Tji. Equation (22) and 
Equation (23) form a pair of dual gravity models. Multiplying equation (22) by equation (23) 
yields 

10 
 
σ
2
2
−
+
+
=
ij
v
u
j
v
u
i
ji
ij
r
P
P
K
T
T
.                            (24) 
From equation (24) it follows equation (1), the power-based gravity model: 
b
ij
j
i
v
u
ij
j
i
v
u
v
u
ji
ij
ij
r
P
P
G
r
P
P
K
T
T
I
=
=
=
+
+
+
)
/
2
)
/(
2
)
/(
1)
(
σ
,                   (25) 
where the gravity coefficient and distance exponent can be expressed as 
)
/(
2
v
u
K
G
+
=
, 
v
u
b
+
=
σ
2
.                             (26) 
This suggests that equation (1) can be derived from the power-based spatial interaction model by 
dint of two allometric relations. According to equation (25) and equations (26), the attractive force 
Iij is a power function of the product of outflow quantity Tij and inflow quantity Tji, and the 
distance exponent b is proportional to the scaling exponent of the spatial interaction model σ. 
Similarly, the exponential-based gravity models can be derived from Wilson’s model. The 
results are as follows 
c
ij r
r
v
j
u
i
ij
e
P
KP
T
/
−
=
,                              (27) 
 
c
ij r
r
u
j
v
i
ji
e
P
KP
T
/
−
=
,                              (28) 
where rc is a spatial scale parameter, which is associated with the average value of the distances 
between n places. Equations (27) and (28) combine to yield 
c
ij r
r
v
u
j
v
u
i
ji
ij
e
P
P
K
T
T
/
2
2
−
+
+
=
,                           (29) 
from which it follows the exponential-based gravity model, equations (6), that is 
0
/
]
)
/[(
2
)
/(
2
)
/(
1)
(
r
r
j
i
r
v
u
r
j
i
v
u
v
u
ji
ij
ij
ij
c
ij
e
P
GP
e
P
P
K
T
T
I
−
+
−
+
+
=
=
=
.             (30) 
The gravity coefficient and spatial scale parameter can be expressed as 
)
/(
2
v
u
K
G
+
=
, 
2
)
(
0
cr
v
u
r
+
=
.                          (31) 
It can be proved that the scale parameter of the exponential-based model, r0, equals half average 
distance. This suggests that the scale parameter can be estimated with the mean of the distances 
between the n places,r , which is defined by 
∑∑
∑∑
=
=
=
=
+
=
+
=
n
i
n
j
ij
n
i
i
j
ij
r
n
n
r
n
n
r
1
1
1
1
)1
(
1
)1
(
2
.                     (32) 
Thus we have r =2r0=(u+r)rc, which will be confirmed by the following empirical evidence. 

11 
 
3 Empirical analysis 
3.1 Materials and methods 
The systems of urban and regional systems of China are employed to verify the distance-decay 
function and the main parameter of the gravity model. The study area involves the entire mainland 
of China. Three measures are adopted, including city size (mass), traffic mileage (distance), and 
freight volume (the function of gravity). The city size is measured with total urban population in a 
region, the traffic mileage is measured with the railway distance between the capital cities of two 
regions, and the freight volume is measured with the quantity of goods carried by railway trains 
from one region to another region. The data of urban population is from the 5th population census 
of China in 2000, the matrix of railway distances of cities is extracted from the mileage table of 
Atlas of China Transportation, and the matrix of railway freight volumes comes from 2001 Year 
Book of China Transportation and Communications. The urban population data were processed by 
the well-known urban geographer, Yixing Zhou, and one of his co-workers. Zhou and Yu (2004a; 
2004b) have published the population data of the 666 China’s cities based on the 5th census data. 
It is easy to sum the urban population of all cities in a region to yield a dataset of regional city 
sizes. There are 31 administrative regions in Mainland China, which comprise 22 provinces, 5 
autonomous regions, and 4 municipalities directly under China’s Central Government. Two special 
administrative districts and a special island of China, Hongkong, Macao, and Taiwan, do not 
belong to Mainland China. The railway network had not reached to the autonomous region of 
Tibet and Hainan province until 2000. So, only 29 regions are taken into account, i.e., the sample 
size is n=29, and the tables of railway distances and freight volumes make two 29×29 matrixes.  
The ordinary least squares (OLS) method is employed to evaluate the parameters of the gravity 
models. The reason for this is that the OLS method bears an advantage in estimating the 
regression coefficients of a linear equation. Other methods such as curvefitting method may get 
more accurate values for the constants G and K, but it often cannot give creditable values of the 
parameters u, v, σ, and rc. In geographical analysis, the slope is more important than the intercept 
in both theoretical studies and empirical analysis. Because of this, the OLS method is more 
suitable for estimating the parameters of the models based on power laws than other algorithms. 
The parameters of the gravity models can be estimated by the log-linear least squares regression. 

12 
 
Taking the logarithms of equations (22), (23), and (24) yields three multivariable linear equations 
ij
j
i
ij
r
P
v
uP
K
T
ln
ln
ln
ln
σ
−
+
+
=
,                        (33) 
ij
j
i
ji
r
P
u
vP
K
T
ln
ln
ln
ln
σ
−
+
+
=
,                        (34) 
ij
j
i
ji
ij
r
P
v
u
P
v
u
K
T
T
ln
2
ln
)
(
ln
)
(
ln
2
)
ln(
σ
−
+
+
+
+
=
.              (35) 
In virtue of these equations, we can estimate the parameters of the power-based models. Similarly, 
taking logs on both sides of equations (27), (28), and (29) gives 
ij
c
j
i
ij
r
r
P
v
uP
K
T
1
ln
ln
ln
−
+
+
=
,                         (36) 
ij
c
j
i
ji
r
r
P
u
vP
K
T
1
ln
ln
ln
−
+
+
=
,                         (37) 
ij
c
j
i
ji
ij
r
r
P
v
u
P
v
u
K
T
T
2
ln
)
(
ln
)
(
ln
2
)
ln(
−
+
+
+
+
=
.                (38) 
By means of the three equations, we can compute the parameter values of the exponential-based 
models using the least squares regression. 
3.2 Results and analysis 
The interregional traffic flow of the 29 Chinese regions falls into two types: outflow and inflow. 
The former is a kind of output flow (Tij), and the latter, input flow (Tji). Using the city sizes, Pi and 
Pj, and the railway distances, rij, as three independent variables, and Tij as a dependent variable, 
we can make a multivariate log-linear regression analysis. Fitting the data to equation (33) yields 
1996
.1
6790
.0
4604
.0
003964
.0
−
=
ij
j
i
ij
r
P
P
T
,                         (39) 
which is the outflow model. The goodness of fit is about R2=0.5227. The distance exponent can be 
estimated with equations (26), which gives b≈2*1.1996/(0.4604+0.6790) ≈2.1056. The 
symmetrical expression of equation (39) can be obtained by fitting the data to equation (34), and 
the inflow model is 
1996
.1
4604
.0
6790
.0
003964
.0
−
=
ij
j
i
ji
r
P
P
T
,                         (40) 
The goodness of fit and the estimated value of the distance exponent based on the second model 
are the same with those based on the first model. A least squares regression of equation (35) yields 

13 
 
3993
.2
1394
.1
1394
.1
000015714
.0
−
=
ij
j
i
ji
ij
r
P
P
T
T
,                      (41) 
which is the integrated model. The goodness of fit is about R2=0.6167. The distance exponent is 
b≈2.3993/1.1394≈2.1056. Obviously, multiplying equation (39) by equation (40) on the same 
sides produces equation (41). In light of equation (25), equation (41) can be transformed into the 
power-based gravity model of Chinese fright flow. Extracting the 1.1394th root of equation (41) 
on both sides yields the normal gravity model as below 
1056
.2
00006082
.0
−
=
ij
j
i
ij
r
P
P
I
.                          (42) 
According to equation (42), the gravity coefficient is about 0.00006082, and distance exponent is 
b≈2.1056 with a standard error sb≈0.083. This is contrast with the result of Rybski et al (2013), 
who once estimated a distance-decay exponent about 2.5 for Paris and its surroundings.  
In order to compare the effect of the power-based gravity model with that of the 
exponential-based model, we should make regression analyses for the gravity model based on 
exponential distance-decay. Using the least squares calculation to fit the data to equations (36), 
(37), and (38) yields 
9268
.
1480
/
6871
.0
4684
.0
000001562
.0
ijr
j
i
ij
e
P
P
T
−
=
,                     (43) 
9268
.
1480
/
4684
.0
6871
.0
000001562
.0
ijr
j
i
ji
e
P
P
T
−
=
,                     (44) 
4634
.
740
/
1554
.1
1554
.1
02441
0000000000
.0
ijr
j
i
ji
ij
e
P
P
T
T
−
=
,                (45) 
in which 1480.9268 is the estimated value of the parameter rc. The values of the goodness of fit 
are 0.4687, 0.4687, and 0.5522, respectively. Actually, equations (43) and (44) combine to make 
equation (45), from which it follows the exponential-based gravity model of Chinese fright flow 
5950
.
855
/
8917
0000000000
.0
ijr
j
i
ij
e
P
P
I
−
=
,                    (46) 
which gives the estimated value of the spatial scale parameter, r0=855.5950. The average distance 
can be estimated as 
*
r =2r0=1711.1900. The mean of the observational distances can be 
computed by using equation (32), which yields r ≈1780.5885. Clearly, the r value is close to the 
2r0 value. This suggests that the scale parameter is associated with the mean of distances between 
different places. The r0 indicates a characteristic length of exponential-based spatial interaction. 
From the results of the regression analysis, two conclusions can be drawn as follows. First, the 
inflow quantity of a region is not equal to its outflow, but there exists symmetrical 

14 
 
relationships between gravity models for inflow and those for outflow. Though the inflow of a 
zone does equal its outflow, the total inflow equals total outflow of all the zones. Because of this, 
the inflow gravity models are symmetric to the outflow gravity models. In practice, we need a pair 
of dual gravity models to describe different flows. Second, the gravity model based on a power 
function is better than the model based on an exponential function where the traffic flows of 
China in 2000 are concerned. The modeling effect can be evaluated by statistic tests. Both the 
global tests and local tests show that the power-based gravity model is more suitable for 
describing the interregional freight flows of china (Tables 1 and 2). 
 
Table 1 Comparison of global effect of the power-based gravity model with that of 
exponential-based gravity model of China’s interregional freight flows (2000) 
Model 
Goodness of fit 
(R2) 
Standard 
error 
Degree of 
freedom 
F 
Statistic 
Significance 
F 
Power-based model 
0.6167 
1.6333 
808 
433.3758 
9.8017E-168
Exponential-based model
0.5522 
1.7655 
808 
332.0667 
1.9172E-140
Note: The standard error suggests the prediction effect of a linear regression model. 
 
Table 2 Comparison of local effect of the power-based gravity model with that of 
exponential-based gravity model of China’s interregional freight flows (2000) 
Model 
Parameter 
Statistic summary 
Multi-collinearity statistic
Item 
Coefficient Standard error
t Statistic
P-value 
Tolerance 
VIF 
Power-based 
model 
Intercept
-11.0612 
1.9939 
-5.5475 
3.9281E-08  
 
lnr 
-2.3993 
0.0946 
-25.3725 
6.4863E-105
0.9687 
1.0323 
lnP1 
1.1394 
0.0758 
15.0312
3.2742E-45
0.9823 
1.0180 
lnP2 
1.1394 
0.0758 
15.0312
3.2742E-45
0.9823 
1.0180 
Exponential 
-based model 
Intercept
-26.7385 
1.9192 
-13.9320
1.0679E-39
 
 
lnr 
-0.0014 
0.0001 
-20.8438 
1.5085E-77
0.9655 
1.0358 
lnP2 
1.15549 
0.0820 
14.0888 
1.8061E-40
0.9806 
1.0198 
lnP1 
1.15549 
0.0820 
14.0888 
1.8061E-40
0.9806 
1.0198 
Note: The standard errors are for evaluating the precision of the linear regression coefficients, VIF means 
“variance inflation factor”. 
 
The derivation of the gravity model from the spatial interaction model is based on an 
assumption that there exist allometric relationships between the size (regional city population) and 
the flow (traffic flow) of the n zones. We can investigate the allometric scaling by fitting equations 

15 
 
(21) to the observational data of China (2000). The first allometry is the power-law relationship 
between total urban population (Pi) and total outflow quantity (Oi). The scatterplot shows that the 
allometric model does not fit the observations very well. The goodness of fit is only R2=0.2743 
(Figure 1(a)). However, three aspects of facts should be noted. First, the scattered points are not 
random distribution. There is a significant trend. Second, among various possible functions 
depicting the trendline, the power function is the most probable one. Third, the significance level 
corresponding to the goodness of fit of the allometric model is about 0.0061, which suggests the 
level of confidence is more than 99.39% (Table 3). The second allometry indicates the scaling 
relationship between total urban population (Pj) and total inflow quantity (Dj). The goodness of fit 
is about R2=0.6127. The power law is the most probable model among various possible relations. 
The significance of fit is about 0.00000052, which implies the confidence level is greater than 
99.99% (Table 3). The effect of the second allometric curve fitting is good in spite of the large 
deviation. It is bias rather than deviation that we must avoid in a statistical analysis. All in all, the 
two allometric models are acceptable according to the general criterion of confidence (significance 
level α=0.05). 
The difference of the fit quality of the two allometric models reflects the state and property of 
China’s economic and urbanization. The goodness of fit of the second model is significantly 
greater than that of the first model Figure 1(b)). This suggests that the urban population of a region 
is mainly correlated with inflow quantity rather than outflow quantity. The inflow reflects the 
inward attraction of a region, while the outflow reflects regional outward action and influence. 
China is a developing country. Both urbanization and industrial development are at the primary or 
intermediate stage (Chen, 2010b). The freight flow comprises principally raw materials such as 
coal, iron ore, and foodstuff, which come directly from earth surface. In this case, there are many 
outliers of allometric relation in the scatterplots. For example, Shanxi province is a less developed 
region in China. Its urbanization level is about 34.91% in 2000. However, Shanxi is the most 
important area of coal producing. The outflow of Shanxi is not proportional to its urban population 
due to coal output. Thus, Shanxi forms a typical outlier, a protrudent point, which is displayed at 
the top left corner of the first subplot of Figure 1(a). The situation of Inner Mongolia (autonomous 
region) is similar to that of Shanxi. Another typical example is Shanghai (municipality), which 
represented the most developed area in Mainland China. The urbanization level of 2000 year is 

16 
 
about 88.31%. However, the freight flow from Shanghai is relatively small so that the outflow is 
not very proportional to its urban population. Thus Shanghai makes another kind of outlier. The 
case of Guangdong province is similar to Shanghai. If we remove the outliers from the dataset, the 
goodness of fit will go up greatly. In fact, geographical laws are the laws of evolution rather than 
the laws of existence. The more developed a geographical system is, the more significant a 
scientific law ruling the system takes on. 
 
Table 3 The goodness of fit and corresponding significance of the possible models for the relation 
between regional city population and freight outflow/inflow of China (2000) 
Relation 
Urban population (Pi) and Outflow (Oi)
Urban population (Pj) and inflow(Dj) 
Testing 
Goodness of fit R2 
Significance 
Goodness of fit R2 
Significance 
2
27
,
01
.0
R
Linear 
0.0449 
0.2698 
0.2753 
0.0035 
0.2214 
Exponential 
0.0916 
0.1105 
0.3323 
0.0011 
0.2214 
Logarithmic 
0.1290 
0.0557 
0.3837 
0.0003 
0.2214 
Power 
0.2473 
0.0061 
0.6127 
0.0000 
0.2214 
Note: The fit of freedom of the regression modeling is n-2=27. If the level of significance is set to α=0.05, the 
threshold of the coefficient of determination is R0.01, 27
2=0.2214. This is to say, if the goodness of fit is greater than 
0.2214, then the level of confidence of the scaling exponent will be greater than 99%. 
 
 
 a. Urban population and outflow                 b. Urban population and inflow 
Figure 1 The allometric scaling relations between total urban population and total freight 
volumes (outflow and inflow) of China’s 29 regions (2000) 
 
O = 0.9547P0.5259
R² = 0.2473
0
5000
10000
15000
20000
25000
30000
0
10000000 20000000 30000000 40000000
Outflow quantity Oi
Urban population Pi
D = 0.0222P0.7677
R² = 0.6127
0
5000
10000
15000
20000
25000
0
10000000 20000000 30000000 40000000
Inflow quantity Dj
Urban population Pj

17 
 
4 Questions and discussion 
The gravity model based on power law proved to be associated with fractal dimension and 
allometric scaling. Assuming that the transport cost is linearly proportional to distance, Wilson 
(1970) derived the exponential-based spatial interaction model using the entropy-maximizing 
method. The transport gravity model based on exponential decay can be inferred from Wilson’s 
spatial interaction model. Postulating that the rate of growth of transport cost with distance is in 
inverse proportion to distance, I derive the power-based spatial interaction from the principle of 
entropy maximization; then postulating allometric relations between inflow/outflow and 
population sizes of different places, I derive the power-based gravity model. The dimension 
problem of the gravity model based on power law can be solved by using the ideas from fractal 
dimension. However, the locality of the spatial action is still a difficult problem for the 
exponential-based gravity model. An inevitable trend of development is a regression of the 
exponential-based gravity model to the gravity model based on power law. 
The power-based gravity model relate to the underlying rationale of theoretical geography. Both 
the first law and the second law of geography are actually based on the original gravity model. 
The first law of geography of Tobler (1970) reads “everything is related to everything else but 
near things are more related than distant things”; the second law of geography of Arbia et al (1996) 
asserts “everything is related to everything else, but things observed at a coarse spatial resolution 
are more related than things observed at a finer resolution”. The first law reflects the inverse 
proportional relation between the gravity and distance, while the second law reflects the direct 
proportional relation between gravity and population size product. Both the geographical laws are 
based on action at a distance. However, the exponential-based distance decay defies the 
long-distance effect but support the localization. The negative exponential bears a characteristic 
scale r0. If the distance is greater than the special scale, i.e., r>r0, the gravity of a place following 
the negative exponential law will attenuate rapidly and then fade out. If the spatial action is 
dominated by an inverse power law, the gravity will tail off but never vanish. In fact, the 
distinction between the power-law decay and exponential decay can be revealed by means of 
autocorrelation function (ACF) and partial autocorrelation function (PACF) analysis. ACF 
includes direct autocorrelation and indirect autocorrelation, while PACF only denotes direct 

18 
 
autocorrelation, no indirect autocorrelation. Both the ACF and PACF of the power-law decay tail 
off gradually (Figures 2(a) and 2(b)). As for the exponential decay, the ACF tails off (Figures 2(c)), 
but the PACF cuts off after the first order displacement (Figures 2(d)) (Chen, 2008; Chen, 2012). 
This suggests that the direct spatial action of a place based on exponential decay is significantly 
localized and cannot reach distance places. The differences and similarities of the two 
distance-decay functions have been previously discussed for many times (Batty and Kim, 1992; 
Chen, 2014b). A simple comparison is drawn between the two types of gravity models (Table 4). 
 
Table 4 Comparison between the exponential-based gravity model and the power-based gravity 
model 
Items 
Exponential-based gravity model 
Power- based gravity model 
Distance-decay function 
Negative exponential function 
Inverse power function 
Dimension 
Euclidean dimension 
Fractal dimension 
Spatial pattern 
Simple network 
Complex network 
Spatial process 
Localization 
Action at a distance 
Spatial memory 
No memory 
Long-range memory 
Symmetry 
Scale translational symmetry 
Scaling symmetry 
Basic principle 
Entropy maximization 
Dual entropy maximization 
 
 
     a. ACF of power-law decay                  b. PACF of power-law decay 
 
-0.2 
-0.1 
0.0 
0.1 
0.2 
0.3 
0.4 
0.5 
0.6 
0.7 
ACF
Displacement 
-0.2 
-0.1 
0.0 
0.1 
0.2 
0.3 
0.4 
0.5 
0.6 
0.7 
ACF
Displacement 
-0.2 
0.0 
0.2 
0.4 
0.6 
0.8 
1.0 
1.2 
ACF
Displacement 
-0.2 
0.0 
0.2 
0.4 
0.6 
0.8 
1.0 
1.2 
ACF
Displacement 

19 
 
    c. ACF of exponential decay                 d. PACF of exponential decay 
Figure 2 The histograms of spatial ACF and PACF based on the inverse power function and 
negative exponential function (by Chen, 2012) 
Note: The two horizontal lines or approximate horizontal lines represent the “two-standard-error bands” of ACF or 
PACF. If the ACF or PACF comes between the two lines, the autocorrelation or partial autocorrelation can be 
omitted at the significance level α=0.05. 
 
One of the defects of this study rests with the assumption of the cost-distance relation. The 
single logarithmic function on the relation between the distance and transport cost has not been 
theoretically proved using mathematical reasoning and empirical verified using observational data. 
The inverse function of a logarithmic function is an exponential function. In principle, the 
logarithmic relation is also derivable using entropy-maximizing method. However, the process of 
derivation remains to be researched in future. In fact, how to make a postulate depends on a 
scientist himself, but the model based on the postulate must be able to explain the observed facts 
and predict the unknown facts (Komuro, 2001). Compared with the exponential-based model, the 
power-based model can better explain the human geographical systems. Both urban and regional 
systems are complex spatial systems (Allen, 1997; Wilson, 2000). It is power laws rather than 
exponential laws that can be employed to interpret spatial complexity of human geographical 
systems (Batty, 2005; Batty and Longley, 1994; Frankhauser, 1994). The empirical case is also 
deficient owing to absence of available data of passenger flow, which results in an incomplete 
component of traffic flow. Despite this, the calculations are revealing for our understanding the 
gravity models. In short, the case study is not best, but we cannot find better one because of 
difficulty of finding observational data for a social scientific study. 
5 Conclusions 
Fractal geometry provides an important mathematical tool for geography and social physics. 
Twenty years ago, Batty (1992) predicted that many of our theories in physical and human 
geography could be reinterpreted using ideas from fractals. Actually, the basic geographical 
models based on power laws such as the allometric growth and gravity models were once 
abandoned because of dimensional dilemma in theory rather than uselessness in practice. Today, 

20 
 
by means of the notions from fractal geometry, we can retrieve these useful power-based models. 
In a sense, fractal theory raises the gravity model based on the power-law distance decay 
phoenix-like from the ashes. From the mathematical derivation and empirical analysis, the main 
conclusions can be drawn as follows. 
First, the gravity model based on power law can be derived from the principle of entropy 
maximization. For the spatial variable from 0 to infinity, entropy-maximization results in negative 
exponential distributions. The exponential-based spatial interaction model and gravity model are 
derivable using the entropy maximizing methods. A power-law distribution can be deduced from a 
pair of exponential distributions. This suggests that a power law is based dual entropy-maximizing 
processes, which are of unity of opposites in self-organized evolution. Replacing the linear 
distance-cost relation with a logarithmic relation between transport cost and distance, we can 
derive the power-based spatial interaction model from entropy-maximization hypothesis of 
regional systems. From the spatial interaction model, we can further derive the gravity model by 
introducing two allometric realtions. The inverse function of a logarithmic function is an 
exponential function, which is theoretically derivable from the ideas of entropy maximization. 
Second, the most proper distance-decay function of the gravity model is the inverse power 
function rather than the negative exponential function. Both the exponential decay distribution 
and power-law decay distribution can be derived from general principles, and have firm 
theoretical basis. The dimension problem of the power-based gravity model can be solved with the 
ideas from fractals. The distance exponent of the gravity model is a fractal dimension of size 
measures, and the scaling exponent of the corresponding spatial interaction model can be 
associated with the fractal dimension of size measures. On the other hand, the locality of the 
exponential-based gravity model cannot be solved at present. The locality is incompatible with the 
action at a distance, and thus does not support spatial interaction. What is more, the power-based 
gravity model can be employed to interpret the complex patterns of geographical systems. 
However, the exponential –based gravity model indicates simplicity rather than spatial complexity. 
Third, the gravity models are different from spatial interaction models, but they can be 
connected with one another by allometric relations. The spatial interaction model can be 
directly derived from the principle of entropy maximization, while the gravity model can be 
derived from the spatial interaction model by way of the allometric relations between population 

21 
 
sizes and inflow/outflow quantities. The gravity model can be used to measure the strength of 
association of one place with another place, while the spatial interaction model can be used to 
analyze the spatial networks and dynamics of flows or migrations. The gravity model states that 
the attractive force between two places is proportional to the product of the two places’ population 
sizes and inversely proportional to the distance between the centroids of two places. In contrast, 
the spatial interaction model states that flow or migration between two places is proportional to 
the product of the inflow quantity and outflow quantity of the two places and inversely 
proportional to the intervening distances. 
Fourth, the human force of attraction between two places is not equal to the interregional 
traffic flow quantity, and we need a pair of dual gravity models to describe inflows and 
outflows. General traffic flow, including inflow and outflow, immigration and emigration, results 
from the gravity of a central place and in turn measures the gravity. A flow quantity is the function 
of gravity in theory, but in technique, the gravity can be treated as a power function of traffic flow 
quantity or a function of the product of inflow and outflow so the human attraction forces are 
measurable in spatial analysis. On the other hand, we need dual gravity models to describe 
interregional inflow and outflow. The relationship between inflow and outflow is asymmetric, but 
there is symmetric relationship between the gravity models for inflow and that for outflow. From 
the dual gravity models it follows the normal gravity models, including the power-based model 
and the exponential-based one. 
 
Acknowledgement: 
This research was supported financially by the National Natural Science Foundation of China (Grant 
no. 41171129).  
 
References 
1. 
Allen PM (1997). Cities and Regions as Self-Organizing Systems: Models of Complexity. 
Amsterdam: Gordon and Breach Science 
2. 
Arbia G, Benedetti R, Espa G (1996). Effects of the MAUP on image classification. Geographical 

22 
 
Systems, 3: 123-141 
3. 
Batty M (1992). The fractal nature of geography. Geographical Magazine, 64(5): 34-36 
4. 
Batty M (2005). Cities and Complexity: Understanding Cities with Cellular Automata, 
Agent-Based Models, and Fractals. Cambridge, MA/London, England: The MIT Press 
5. 
Batty M (2008). The size, scale, and shape of cities. Science, 319: 769-771 
6. 
Batty M, Kim KS (1992). Form follows function: reformulating urban population density 
functions. Urban Studies, 29(7):1043-1070 
7. 
Batty M, Longley PA (1994). Fractal Cities: A Geometry of Form and Function. London: 
Academic Press 
8. 
Batty M. and Karmeshu (1983). A strategy for generating and testing models of migration and 
urban growth. Regional Studies, 17(4): 223-236 
9. 
Bertalanffy L von (1968). General System Theory: Foundations, Development, and Applications. 
New York: George Braziller (4th printing, 1973) 
10. Carey HC (1858). Principles of Social Science, Philadelphia: J.B. Lippincott 
11. Casti JL (1996). Would-Be Worlds: How Simulation Is Changing the Frontiers of Science. New 
York: John Wiley and Sons 
12. Chen YG (2008). A wave-spectrum analysis of urban population density: entropy, fractal, and 
spatial localization. Discrete Dynamics in Nature and Society, Volume 2008, Article ID 728420, 
22 pages 
13. Chen YG (2009). Urban gravity model based on cross-correlation function and Fourier analyses 
of spatio-temporal process. Chaos, Solitons & Fractals, 41(2): 603-614 
14. Chen YG (2010a). Characterizing growth and form of fractal cities with allometric scaling 
exponents. Discrete Dynamics in Nature and Society, Volume 2010, Article ID 194715, 22 pages 
15. Chen YG (2010b). The corresponding relationships between the demographic transition, 
urbanization process, and industrial development of China. Geographical Research, 29 (12): 
2109-2120 [In Chinese] 
16. Chen YG (2012). On the four types of weight functions for spatial contiguity matrix. Letters in 
Spatial and Resource Sciences, 5(2): 65-72 
17. Chen YG (2014a). An allometric scaling relation based on logistic growth of cities. Chaos, 
Solitons & Fractals, 65: 65-77 

23 
 
18. Chen YG (2014b). Power-law distributions based on exponential distributions: Latent scaling, 
spurious Zipf’s law, and fractal rabbits. Fractals, 22(x): 1400054 
19. Chen YG, Jiang SG (2009). An analytical process of the spatio-temporal evolution of urban 
systems based on allometric and fractal ideas. Chaos, Solitons & Fractals, 39(1): 49-64 
20. Converse PD (1949). New laws of retail gravitation. Journal of Marketing, 14(3): 379- 384 
21. Erlander S (1980). Optimal Spatial Interaction and the Gravity Model. Berlin/New York: 
Springer-Verlag 
22. Frankhauser P (1994). La Fractalité des Structures Urbaines (The Fractal Aspects of Urban 
Structures). Paris: Economica (In French) 
23. Gould PR (1972). Pedagogic review: entropy in urban and regional modelling. Annals of the 
Association of American Geographers, 62(1): 689-700 
24. Grigg DB (1977). E.G. Ravenstein and the ‘laws of migration’. Journal of Historical Geography, 
3(1): 41-54 
25. Haggett P (2001). Geography: a Global Synthesis. New York: Pearson Hall 
26. Haggett P, Cliff AD, Frey A (1977). Locational Analysis in Human Geography (2nd edition). 
London: Arnold 
27. Haynes AH (1975). Dimensional analysis: some applications in human geography. Geographical 
Analysis, 7(1): 51-68 
28. Haynes KE, Fotheringham AS (1984). Gravity and Spatial Interaction Models. London: SAGE 
Publications 
29. Komuro N (2001). Suugaku Kiraina Hito No Tame No Suugaku. Tokyo: Toyo Keizai (In 
Japanese) 
30. Mackay JR (1959). The interactance hypothesis and boundaries in Canada: a preliminary study. 
The Canadian Geographer, 3(11):1-8 
31. Mandelbrot BB (1983). The Fractal Geometry of Nature. New York: W. H. Freeman and 
Company 
32. Ravenstein EG (1885). The laws of migration. Journal of the Statistical Society of London, 48(2): 
167-235 
33. Reilly WJ (1931). The Law of Retail Gravitation. New York: The Knickerbocker Press 
34. Rybski D, Ros AGC, Kropp JP (2013). Distance-weighted city growth. Physical Review E, 87(4): 

24 
 
042114 
35. Sen A, Smith TE (1995). Gravity Models of Spatial Interaction Behavior: Advances in Spatial and 
Network Economics. Berlin/New York: Springer Verlag 
36. Taylor PJ (1977). Quantitative Methods in Geography. Prospect Heights, Illinois: Waveland Press 
(3th printing,1983) 
37. Tobler W (1970). A computer movie simulating urban growth in the Detroit region. Economic 
Geography, 46(2): 234-240 
38. Tobler W (2004). On the first law of geography: A reply. Annals of the Association of American 
Geographers, 94(2): 304-310 
39. Wilson AG (1970). Entropy in Urban and Regional Modelling. London: Pion Press 
40. Wilson AG (2000). Complex Spatial Systems: The Modelling Foundations of Urban and Regional 
Analysis. Singapore: Pearson Education 
41. Zhou YX, Yu HB (2004a). Reconstruction city population size hierarchy of China based on the 
fifth population census (I). Urban Planning, 28(6): 49-55 [In Chinese] 
42. Zhou YX, Yu HB (2004b). Reconstruction city population size hierarchy of China based on the 
fifth population census (II). Urban Planning, 28(8): 38-42 [In Chinese] 
