Journal of Mathematical Economics 67 (2016) 125–137

Contents lists available at ScienceDirect

Journal of Mathematical Economics

journal homepage: www.elsevier.com/locate/jmateco

Spatial growth with exogenous saving rates✩

A. Xepapadeas, A.N. Yannacopoulos ∗
Athens University of Economics and Business, Greece

article info abstract

Article history: Economic growth has traditionally been analyzed in the temporal domain, while the spatial dimension
Received 21 December 2015 is captured by cross-country income differences. Data suggest great inequality in income per capita
Received in revised form across countries, and a slight but noticeable increase in inequality across nations between 1960 and
20 July 2016
2000. Seeking to explore the mechanism underlying the temporal evolution of the cross sectional
Accepted 28 September 2016
Available online 14 October 2016
distribution of economies, we develop a spatial growth model where saving rates are exogenous. Capital
movements across locations are governed by a mechanism under which capital moves toward locations
Keywords:
of relatively higher marginal productivity, with a velocity determined by the existing stock of capital. This
Spatial growth augments the capital accumulation equation by a nonlinear diffusion term. Our results suggest that under
Nonlinear diffusion diminishing returns, the growth process leads to a stable spatially nonhomogeneous distribution for per
Solow model capita capital and income in the long run. Insufficient savings may lead to the emergence of persistent
poverty cores where capital stock is depleted in some locations.
© 2016 Elsevier B.V. All rights reserved.

1. Introduction at different steady states may not, however, adequately describe

Economic growth, in formal growth models, has traditionally
been analyzed in the temporal domain with the main focus of
analysis being the development of models capable of explaining
stylized facts, which are expressed in terms of the temporal
evolution of key variables such as output or capital per capita or
the capital–labor ratio. A central issue, however, is cross-country
income differences, which exemplifies the spatial dimension
of the problem. Acemoglu (2009) (Chapter 1), points out that
there is great inequality in income per capita and income per
worker across countries, and that there is a slight but noticeable
increase in inequality across nations between 1960 and 2000. The
geographical or spatial dimension is also taken into account in the
context of convergence. Data suggest (e.g. Acemoglu, 2009) that
there is no unconditional convergence during the post-war period.
However, the results of Barro and Martin (2004) suggest that
conditional convergence takes place with poor countries growing
faster in terms of per capita GDP than rich ones, within a group
that shares similar characteristics. Conditional convergence even
the evolution of the spatial distribution of per capita GDP across
countries. As Quah (1996) (p. 1053) points out, ‘‘Convergence
concerns poor economies catching up with rich ones. What one
wants to know here is, what happens to the entire cross sectional
distribution of economies, not whether a single economy is
tending toward its own, individual steady state.’’ Bourguignon and
Morrisson (2002) (p. 742) report that ‘‘ . . . world income inequality
worsened dramatically over the past two centuries. The Gini
coefficient increased 30% and the Theil index 60% between 1820
and 1992. This evolution was due mainly to a dramatic increase in
inequality across countries or regions of the world.’’ Similar results
have also been reported more recently by Milanovic (2009).
Some insights into the characteristics of the spatial distribution
of GDP per capita can be obtained by using the quantities
  yit − yjt 2
DSt = ,
i̸=j
ȳ
  yit − yjt 
DAt = 
 ȳ 
 j = 1, . . . , N , t = 1980, . . . , 2011
i̸=j

where yit , yjt denote per capita GDP in countries i, j at time t for
a sample of countries i, j = 1, . . . , N, and ȳ denotes the overall
✩ We would like to thank William Brock for many challenging ideas and
suggestions, and an anonymous referee for valuable comments. This research has
average (for all countries) per capita GDP. These quantities can
been co-financed by the European Social Fund – and Greek national funds through
the Program: ARISTEIA – AUEB, ‘‘Spatiotemporal Dynamics in Economics’’. be regarded as providing a measure of spatial nonhomogeneity
∗ Corresponding author. of GDP per capita, in the sense that an increasing measure over
E-mail address: ayannaco@aueb.gr (A.N. Yannacopoulos). time means that the spatial distribution of GDP becomes more
http://dx.doi.org/10.1016/j.jmateco.2016.09.010
0304-4068/© 2016 Elsevier B.V. All rights reserved.
126 250 A. Xepapadeas, A.N. Yannacopoulos / Journal of Mathematical Economics 67 (2016) 125–137

DS
2010). Models of optimal development over space and time, which
DA
Linear [DS]
could be regarded as a suitable vehicle for studying economic
Linear [DA] growth in a geographical context, were developed in the 1970s by
200

Isard and Liossatos (see e.g., Isard and Liossatos, 1973; Isard and
Liossatos, 1973; Isard and Liossatos, 1978; Carlson et al., 1991).
Dynamic spatial economic modeling was developed in the context
150
DS,DA

of economic growth and resource management mainly during the

2000s (e.g., Brito, 2004; Camacho and Zou, 2004; Boucekkine et al.,
2009; Boucekkine et al., 2013a,b; Brock and Xepapadeas, 2008,
100

2010; Brock et al., 2014a,b). The main feature of current spatial

growth models is that the spatial movements of the stock of capital
50

across locations are modeled through a trade balance approach

with respect to a closed region where capital flows are such that
capital is received from the left of the region and flows away to the
0

1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011
right of the region. This leads to a model of classic local diffusion
Year
with a constant diffusion coefficient. Modeling capital movements
Fig. 1. Regional nonhomogeneity measures. in this way implies that capital stock moves from locations of high
capital concentration (i.e. rich countries) to locations of low capital
spatially heterogeneous or ‘‘less flat’’ relative to space.1 Thus concentration (i.e. poor countries). This property seems not to be
increasing measures over time indicate that the dispersal of per compatible with empirical findings since – as indicated by Lucas
capita GDP across the countries in the sample increased during in the context of the Lucas paradox (see, e.g., Lucas, 1990; Lucas,
the sample period. The two nonhomogeneity measures, along with 2003; Prasad et al., 2007) – capital does not seem to flow from rich
the corresponding linear trend, are presented in Fig. 1 for eleven countries to poor countries.
regions of the world,2 , 3 for the period 1980–2011. In the present paper, we contribute to the ongoing research
The evolution of both measures has a very similar pattern, and on spatiotemporal dynamics and spatial growth by developing a
the associated linear trend suggests that the overall dispersal is model in which the basic mechanism underlying the movements
increasing at the regional level.4 This observation, although broad of capital across space is the quest for locations where the marginal
in nature, indicates that the spatial distribution of GDP per capita productivity of capital is relatively higher than the productivity at
does not tend to become more uniform with the passage of time or, the location of origin, without imposing the constraint that capital
to put it differently, does not seem to converge to a geographical moves from locations of high concentration to locations of low
homogeneous state for countries grouped in the traditional way concentration. By assuming that capital flows toward locations of
according to their level of per capita GDP. Countries that start high returns, which is a plausible assumption underlying capital
with lower per capita income in the region may grow faster than flows with endogenous velocity depending on the existing stock
high income countries, which is consistent with β convergence of capital, our model implies that the spatiotemporal evolution
arguments, but this growth does not seem to result in a spatially of capital is governed by a nonlinear diffusion equation. In this
flatter distribution in the long run. case, the ‘‘diffusion coefficient’’ is not constant but depends on
In this context, the purpose of this paper is to develop a spa- the capital stock and the rate of change of marginal productivity
tial model of economic growth and, by doing so, to explore mech- of capital (the second derivative of the production function). This
anisms that could generate, through economic forces, persistent approach for modeling capital flows essentially differs from the
nonuniform spatial distributions of per capita capital and GDP classic diffusion models used in the existing literature which are
across locations, and determine the temporal evolution of these based on the trade balance (e.g., Carlson et al., 1991; Brito, 2004;
spatial distributions. In a sense, we are exploring how traditional Camacho and Zou, 2004; Boucekkine et al., 2009; Boucekkine et al.,
neoclassical growth theory can be extended to a spatial growth 2013a,b), and describe the spatiotemporal evolution of capital
theory which would provide models capable of approximating per- by a parabolic partial differential equation (PDE) with constant
sistent spatial heterogeneity across countries in terms of per capita diffusion coefficient.
GDP. Our contribution, by using the plausible mechanism in which
Economic geography and economic growth has been discussed capital moves toward locations of higher productivity and not a
in the so-called second generation of new economic geography mechanism in which capital moves necessarily from higher to
models, but not in a formal growth context (e.g., Martin and lower concentrations, is more in line with the mainstream neoclas-
Ottaviano, 2001; Baldwin et al., 2001, 2003; Baldwin and Martin, sical models than the existing spatial growth models. Using stan-
2004; Fujita and Mori, 2005; Desmet and Rossi-Hansberg, 2009, dard neoclassical growth assumptions, this mechanism generates
spatial distributions for per capita capital and GDP which are char-
acterized by large and persistent spatial nonhomogeneities. These
nonhomogeneities could be regarded as compatible with existing
1 The DS measure can be related to a discretized version of a Sobolev norm,
t observations. The notion of capital we employ is a ‘‘mechanistic’’
while the DAt measure, by using the absolute value of pairwise comparisons, can
kind that cannot move very fast, as financial capital does, to ar-
be regarded as consistent with the Lorentz criterion (Ray, 1998).
2 The World Bank database was used for the GDP per capita (PPP constant 2005 eas of high marginal productivity because of adjustment costs and
international $). The regions according to the World Bank classification are: Arab other potential institutional barriers in this location.
World, ARB; Caribbean Small States, CSS; East Asia & Pacific (developing only), EAP; By considering a distance metric concept based on economic
European Union, EU; Europe & Central Asia (developing only), ECA; Latin America distance, we develop local models of capital diffusion and we
& Caribbean (developing only), LAC; Middle East & North Africa (developing only), develop an analytical framework that extends the standard Solow
MNA; North America, NAC; Pacific Island Small States, PSS; South Asia, SAS; Sub-
Saharan Africa (developing only), SSA.
model in a geographical context. The spatial Solow model with
3 We would like to thank Petros Xepapadeas for computational assistance with a mechanism underlying capital flows which leads to nonlinear
the nonhomogeneity measures. diffusion, generates solutions in which spatially nonhomogeneous
4 A similar pattern is detected for the groups of high and low income countries distributions of per capita capital and income across locations
defined according to the World Bank classification. persist over time. In certain cases, spatial nonhomogeneity may
 A. Xepapadeas, A.N. Yannacopoulos / Journal of Mathematical Economics 67 (2016) 125–137
be amplified over time and locations may end up at a steady state
in poverty cores with capital stock approaching zero. Our results
about persistent spatial heterogeneity and nonsmoothing of spatial
differences do not require increasing returns and are obtained
under standard diminishing returns to capital.
2. Modeling the spatiotemporal evolution of capital
An issue that should be addressed by a spatial growth model is
the definition of an appropriate distance metric. The most common
metric of the distance between two spatial points (say countries)
is geographical distance, as measured for example by the distance
between capital cities. Conley and Ligon (2002) suggest that a
more appropriate metric for measuring distances associated with
economic activities is that of the economic distance – the economic
metric – reflected in transportation costs. It turns out that the
distance between countries might be very different, depending on
whether the geographic or the economic metric is used.5
Using the concept of economic distance, we develop a local
model that enables us to study the spatiotemporal evolution of
capital in the context of an economic metric. Since each element
of the economic space can be mapped to one and only one
element of the geographical space, any spatial distribution defined
in economic space can be transformed into a corresponding
distribution in the geographical space. This equivalence allows us
to work with local models defined in the economic space through
local transport operators, an approach which is not appropriate
when capital flows are defined in the geographical space.
We assume that for any (fixed) time instance t, economic
activity takes place in a domain U ⊂ Rd , of sufficiently smooth
boundary Γ = ∂ U with point z ∈ U. The domain U may represent
either physical space (in which case d = 1, 2) or more abstract
economic structures (in which case d can be higher than 2). The
spatiotemporal allocation of per capita capital is then modeled by
a function k : [0, T ] × U → R+ , describing the distribution of
capital in space over time, so that k(t , z ) denotes the density of per
capita capital at time t ∈ [0, T ] at point z ∈ U.
At any time t and spatial location z, we assume that per capita
output, y(t , z ), depends on the per capita capital stock, k(t , z ),
though a neoclassical production function f : R+ → R+ ,
y(t , z ) = A(z )f (k(t , z )),
(1)
which is assumed to be twice differentiable on (0, ∞) and
concave. The factor A accounts for spatial heterogeneities related to
productivity. Our implicit assumption is that labor is immobile.6 A
location z is characterized by an exogenous saving ratio 0 < s < 1
and exponential capital depreciation at a given rate δ > 0.
Capital stock accumulated at location z can be moved out
of z to locations z ′ , or flow into z from locations z ′′ through a
transport mechanism. The transport mechanism prescribes that
capital moves toward locations where its marginal productivity,
m, is higher than that in the location of origin. In order to build
a local spatiotemporal model for k, we need to consider the local
balance equation for capital stock in any region V ⊂ U such that
V ∩ Γ = ∅. This should take into account the balance between
the temporal rate of change of k, the local rate of accumulation of k
through savings, and the net inflow and outflow rate into and out
5 For example, while the geographical distance between Australia and Egypt
is smaller than the distance Australia–UK and Australia–USA, the corresponding
economic distance between Australia and Egypt, both in terms of UPS and airfare,
is larger than the distance Australia–UK and Australia–USA.
6 This assumption may be revisited by including labor mobility in the model,
although it seems justified taking into account the relative difficulty of labor
mobility as compared to capital mobility.
127
of V , and then pass to the limit as V contracts to a single point z. The
rate of net inflow and outflow of capital from V is modeled through
a vector field v : [0, T ] × V → Rd , such that
J (t , z ) := v(t , z )k(t , z ) = (v1 (t , z )k(t , z ), . . . , vd (t , z )) (2)
provides the velocity (flux) of capital stock at location z ∈ V at time
t, with vi (t , z ) corresponding to the component of capital velocity
in direction i ∈ {1, . . . , d}. The ansatz (2) reflects the fact that
only a fraction of the capital at location t will relocate so that, in a
sense, vi (t , z ) can be interpreted as the propensity of capital stock
at location z at time t to move along direction i.
Assumption 1. At any point (t , z ) ∈ [0, T ] × U, the velocity field
v depends on the spatial gradients of the marginal productivity
∂y
of capital m(t , z ) = ∂ k (t , z ) and the stock of capital k(t , z )
accumulated at z, as follows:
v(t , z ) = B(z )ψ(k(t , z ))∇z m(t , z ),
where ψ : R+ → R+ is a function reflecting the assumption that
the tendency of capital to move from location z may depend on
the existing stock of capital at z, and B : U → R+ is a function
modeling specific location characteristics.
Assumption 1 models the fact that, at any location z, capital
stock will try to relocate toward the direction where marginal
productivity of capital, m, increases, and the (local) direction of
maximum  increase of m is captured by the gradient vector field
∂m
∇z m = ∂ z1
, . . . , ∂∂zmd . If ψ is an increasing function of k, then
an increase in the stock of capital at z increases the tendency of
capital to move to another location, provided that it can attain
a higher marginal productivity at the new location. The opposite
holds if ψ is a decreasing function.7 Using the above assumptions
and the Gauss divergence theorem, we may obtain a local model
for the spatiotemporal evolution of capital concentration in terms
of a nonlinear PDE.
Proposition 1. If capital movements follow Assumption 1, then the
spatiotemporal evolution of the capital stock, when saving rates s (z )
are exogenous across locations, is given in terms of the solution of the
nonlinear diffusion equation
∂k f ′ (k) ∇ A
  
= −div ABψ(k)kf ′′ (k) ∇ k + ′′
∂t f (k) A
+ sAf (k) − δ k, (3)
which is assumed to hold for any (t , z ) ∈ (0, T ] × int(U), and
d ∂
div = j =1 ∂ z i .
If A does not depend on space, then Eq. (3) reduces to the simpler
equation
∂k
= −div ABψ(k)kf ′′ (k)∇ k + sAf (k) − δ k.
 
(4)
∂t
Proof. For the proof, see Appendix A.1 in Appendix. 
For the behavior of k on the boundary points, z ∈ ∂ U, we
assume Neumann boundary conditions J (t , z ) · n = 0, where
n is the outward normal at ∂ U, corresponding to no flux of the
capital stock from the boundary of U. Alternative assumptions
could involve Dirichlet boundary conditions k(t , z ) = 0 for z ∈
∂ U, periodic boundary conditions, or treating U as an infinite
7 The existence and the monotonicity of ψ is an empirical issue. Our modeling
framework is quite general and capable of incorporating alternative assumptions.
If, for example, the stock of capital does not affect the flux, then ψ = 1.
128 A. Xepapadeas, A.N. Yannacopoulos / Journal of Mathematical Economics 67 (2016) 125–137
domain and ∂∂z k(t , z ) → 0 as |z | → ∞. Periodic boundary
conditions lead to the same results as Neumann conditions.
Furthermore, Neumann and periodic boundary conditions allow
for the endogenous emergence of spatial patterns and do not
induce them exogenously as the Dirichlet conditions do.
Eq. (3) is a nonlinear diffusion equation in which the transport
mechanism is nonlinear and depends on the underlying production
function. This can be contrasted with spatial growth models based
on the trade balance approach which assume a linear diffusion
transport mechanism and usually admit the form of a semilinear
equation
∂k
= Ddiv∇ k + sAf (k) − δ k.
∂t
The nonlinear form of the transport mechanism in our model (3)
stems from the assumption that capital moves toward locations
of higher marginal productivity, which underlies the structure of
capital flows across locations. Some remarks are in order here.
1. Since the production function is concave (decreasing returns),
f ′′ < 0, Eq. (3) is a nonlinear diffusion equation with a positive
nonlinear diffusion coefficient, which leads to a well-posed
parabolic problem.8
2. An alternative way to derive (3) is to consider it as the
continuous limit of a random walk on a discrete lattice, in
which capital stock at a lattice site i may move to any of its
neighboring lattice sites j with probability proportional to the
difference of marginal productivity between site i and site j. If
m(t , j) > m(t , i), then the capital stock at i will move to j but
not otherwise. The continuous limit of this random walk will
lead to a PDE similar to (3).
3. A spatial Solow model
We turn now to a detailed study of the implications of capital
flows across locations modeled by (3) for the traditional Solow
model with a Cobb–Douglas production function f (k) = Akα ,
α ∈ (0, 1), where for simplicity A is assumed to be independent
of z.
Assumption 2. The tendency of capital to move across locations
follows the form in Assumption 1 with ψ(k) = kρ and (for
simplicity) with B independent of z. If ρ > 0, then an increase
in the capital stock will enforce the tendency of capital to move in
search of higher marginal returns, whereas if ρ < 0, the opposite
will take place. The quantities s and δ may depend on z.
The spatial Solow model is defined in the following proposition,
whose proof is straightforward and thus omitted.
Proposition 2. Under Assumption 2, the fundamental equation of
economic growth describing the spatiotemporal evolution of the
capital stock is given by the quasilinear degenerate PDE
∂ produces phenomena which are never to be observed for linear
k = D̄D0 ∇z · kβ ∇z k + sAkα − δ k,
 
∂t
where D̄ = α(1 − α), β = ρ − (1 − α) and D0 = BA.
8 If we allowed for increasing returns, then the diffusion coefficient would be
negative, leading to an ill-posed parabolic problem. By ill-posed we mean that
even though solutions may exist, we may lose continuity with respect to the initial
condition or have sharp concentrations of capital, blow up phenomena, etc. The case
of increasing returns is an interesting issue for future research.
Except for the special case where β = 0, our model is a
nonlinear diffusion model9 with an effective diffusion coefficient
Deff (t , z ) depending on the state of the system as follows:
Deff (t , z ) := D(k (t , z )) = D̄D0 k (t , z )β , (7)
where D0 is a constant. Therefore, in our model the diffusion
coefficient Deff (t , z ), which determines capital mobility across
space, is determined endogenously under the assumption that
capital flows to locations of relatively higher productivity, unlike
the standard models of linear diffusion in which the fixed diffusion
coefficient D is determined exogenously. We feel that, although the
degree of dependency of the diffusion coefficient on the stock of
(5) capital and the structure of capital velocity is an empirical issue,
our approach – by relating these factors to capital flows – provides
a richer environment for studying the spatiotemporal evolution
of capital stock and could provide a starting point for empirical
research. The qualitative aspects of the spatiotemporal behavior of
capital stock depend on the exponent β in (7). To see this (assuming
for simplicity that D0 = 1), we can express (6) in an equivalent
form as
∂
k = D̄kβ 1k + v · ∇z k + sAkα − δ k, (8)
∂t
where v = β D̄kβ−1 ∇z k is an advection field resulting from the
fact that the diffusion coefficient is nonlinear. (Note that this term
vanishes if β = 0, i.e. when diffusion is linear.) While the Laplacian
term (first term on the right hand side of (8)) corresponds to a
random walk type of motion with equal probability of movement
in each direction, the second term corresponds to a motion of
capital along the direction indicated by the velocity vector −v ,
which is concentration dependent. The direction of −v points
toward the regions where k is decreasing, thus encouraging the
flow of capital toward regions of low capital stock concentration.
However, the intensity or magnitude with which this phenomenon
occurs depends on the capital stock concentration at the particular
region we consider. For example, if 0 < β < 1, then the
magnitude of this phenomenon is almost zero for regions where
k is large (since then kβ−1 → 0), whereas on the contrary it
is very large for regions where k is small (since then kβ−1 →
∞). This means that while large capital concentrations have the
tendency to move toward regions where capital concentration is
small, this movement takes place with a velocity whose magnitude
is decreasing with capital concentration, thus impeding movement
toward locations with small capital concentration. On the contrary,
the opposite phenomenon happens for regions of small capital
concentration, encouraging thus capital outflow from such regions.
Such phenomena are not present in the linear diffusion case where
β = 0. Note that β could potentially be specified from economic
quantities such as the marginal elasticity of production α and
tendency of capital to migrate ρ . The intuition developed in this
section will be verified and enhanced by detailed analytic and
numerical treatment of the model in the remainder of the paper.
In conclusion, we must note that our model, in the absence of the
capital generation term sAf (k) − δ k, reduces to the well-known
and widely studied porous medium equation, which for β > 0
(6)
diffusion, such as for example the finite velocity of propagation
and the sustainability of sharp spatial gradients (see e.g. Vázquez,
2006). It is interesting to note that we arrived at this nonlinear
diffusion model by using economic principles and not by just
overlaying a nonlinear diffusion mechanism on top of a temporal
economic growth model.
9 The special case of an AK model where α = 1, ρ = 0 leads to D̄ = 0.
 A. Xepapadeas, A.N. Yannacopoulos / Journal of Mathematical Economics 67 (2016) 125–137
4. Qualitative and quantitative aspects of the spatial Solow
model
In this section we study the nonlinear spatial growth PDE (6)
characterizing the spatial Solow model defined in Proposition 2.
By an appropriate rescaling of the spatial variable z, the model can
be written as:
∂ variability of the coefficients c1 , c2 . Clearly this PDE always has the
k = D1kβ+1 + sAkα − δ k, (9)
∂t
where D > 0 is a coefficient (independent of k), s is the savings
ratio (possibly spatially varying), A is a productivity parameter, δ
(possibly spatially varying) is the rate of capital depreciation, and
α ∈ (0, 1) is the production elasticity. Furthermore, without loss
of generality, by a rescaling of the variable t, we may express the
above equation in the form
∂
k = 1kβ+1 + c1 kα − c2 k
∂t
where
sA δ c̄1 = sup c1 (z ),
c1 = ≥ 0, c2 = > 0, β = ρ − (1 − α).
D D
Note that in general c1 and c2 could depend on the spatial location
z. We further allow for regions where c1 = 0 (i.e. regions where no
saving is possible). The possibility of allowing for a set U0 ⊂ U ⊂
Rd , with the property c1 (z ) = 0 if z ∈ U0 , may provide insights
regarding the existence of poverty traps.
The PDE (10) of the spatial Solow model will be complemented
with an initial condition k(0, z ) = k0 (z ), where k0 : U ⊂ Rd →
R+ is an initial capital stock distribution, and also with boundary
conditions related to the prescribed behavior of the distribution of
capital stock at certain parts of the domain U. We consider the case
where there is no flux of capital at the boundary, i.e., homogeneous
Neumann boundary conditions n · ∇ kβ+1 = 0 on ∂ U where n is the
outward normal vector on ∂ U. Our final model will therefore be
∂k
= 1Φ (k) + f (z , k), (t , z ) ∈ (U )T ,
∂t
∂ Φ (k)
(t , z ) = 0, (t , z ) ∈ (∂ U )T ,
∂n
k(0, z ) = k0 (z ), z ∈ U ,
where (U )T = [0, T ] × U, (∂ U )T = [0, T ] × ∂ U, and where
Φ : [0, ∞) → [0, ∞), f : U × [0, ∞) → [0, ∞) is defined as:
ρ+α α
Φ (k) := k 1+β
=k , f (z , k) := c1 (z )k − c2 (z )k.
4.1. Steady-state solutions
The starting point for our analysis will be the steady-state
solutions of model (11), i.e., solutions which depend only on z and
not on t. A steady-state solution k∗ = k∗ (z ) can be regarded as a
steady-state distribution of the capital stock across space. For such
solutions, the spatial Solow equation (11) simplifies to:
−1kβ+1 = c1 kα − c2 k, z ∈ U
n · ∇ kβ+1 = 0, z ∈ ∂ U ,
where β + 1 = ρ + α .
In the special case where c1 and c2 are constants, it is easy to see
 1/(1−α)
that (13) admits two possible solutions, k = 0 and k =
which will be called hereafter the flat solutions. These are the
standard solutions of the non-spatial Solow model. However, the
PDE (13) has a rich behavior, and can generate a large variety of
spatially varying steady-state solutions.
129
While the exact spatial distribution of capital may depend
either on the boundary conditions or the specific spatial variability
of the coefficients, there are certain qualitative features which are
robust with respect to these two aspects and in our view offer
interesting insight into the problem. These are collected in the
following propositions.
Our first result is an existence and uniqueness result for the
steady-state PDE (13) in any spatial dimension allowing for spatial
solution k(z ) = 0 for every z ∈ U, which will hereafter be called
the trivial solution.
Proposition 3. Let β ≥ 0, α < 1 and assume Hölder
continuity properties10 for the coefficients c1 , c2 . The steady-state
equation (13) has a classical (nontrivial) positive solution,11 satisfying
the a priori bounds
1
  1−α
c̄1
(10) 0≤k≤ ,
c2
where
c 2 = inf c2 (z ).
z ∈U
z ∈U
The solution is unique if α − 1 < β . The results of the proposition
remain true if we consider Dirichlet or periodic boundary conditions.
Proof. For the proof, see Appendix A.2 in Appendix. 
The spatial model with nonlinear diffusion allows us to explore
cases in which the spatial domain contains locations where
savings vanish. This might be a realistic situation for extremely
impoverished locations. We will call these regions poverty cores
and will define them as regions V0 ⊂ U with the property that
c1 (z ) = 0 if z ∈ V0 . The poverty core suggests the existence of
regions where capital is identically zero at a steady state, implying
that the steady-state distribution of the capital stock contains
regions with no capital and regions with positive capital. Thus
economies where savings are not possible could eventually be
trapped in the poverty core where their capital stock is depleted.
Since poverty cores will coexist with regions of positive capital,
convergence – in the sense used in growth theory – is not possible
(11)
at the steady state. The existence of poverty cores is verified by
the existence of compact support solutions for the steady-state
equation (13) and is established in the following proposition.
Proposition 4. Let β > 0, assume that c1 , c2 are Hölder continuous
functions and that for some z0 ∈ U and ρ > 0, c1 vanishes, due to zero
(12) savings ratio s, inside a ball centered at z0 with radius 2ϱ, situated
in the interior of the spatial domain. Then, any nontrivial positive
solution of the steady-state equation (13) will develop a poverty core,
i.e. a region of total depletion of capital stock inside a ball centered at
z0 and of radius ϱ, as long as the parameters of the problem satisfy the
condition
 − γ1−γ2
c̄1 −γ2 1
− c2 + νϱ2 ((ν − 1) + (d − 1)) ≤ 0, (14)
c2
(13) where ν = 1−γ2
and d is the spatial dimension. Poverty traps cannot
2
form if β = 0. The proposition also holds true for the case of Dirichlet
boundary conditions or periodic boundary conditions.
c1 10 A function g : Rd → R is called Hölder continuous when there exist α < 1 and
c2
,
C > 0 such that |g (x) − g (y)| ≤ C |x − y|α for every x, y ∈ Rd . Hölder continuity is
a form of uniform continuity which is weaker than Lipschitz continuity.
11 The regularity of the solution depends on the regularity of the coefficients c , c .
1 2
If c1 , c2 ∈ L∞ (U ), then the solution is a weak solution k ∈ L1 (U ), whereas if the
coefficients c1 , c2 enjoy Hölder continuity properties, the solution is classical.
130 A. Xepapadeas, A.N. Yannacopoulos / Journal of Mathematical Economics 67 (2016) 125–137

10
Proof. The proof is given in Appendix A.3 in Appendix. 

Poverty cores do not emerge in models of linear diffusion 9

(β = 0) where capital moves from high to low abundance
locations, even if c1 vanishes in subsets of U. Thus nonlinear 8
diffusion (β > 0) can help model the emergence of poverty
cores where capital is depleted due to zero savings. This happens 7
because although capital moves to locations with low capital
stock, since these locations are characterized by high marginal 6

k
productivity of capital as capital is depleted, no part of the inflow
is used for local capital accumulation since nothing is saved. If 5
the parameters are such that condition (14) is satisfied, then no
accumulation will take place at this location and eventually the 4
capital stock will be depleted. The vanishing of savings at a point
is not, however, enough to guarantee the existence of a poverty 3

core. Due to the spatial interactions, a poverty core emerges when

relation (14) is satisfied. This relation links the maximum value of 2
–4 –3 –2 –1 0 1 2 3 4
c1 over a wider region (reflecting saving rates and productivity in z

nearby regions) with the minimum value of c2 over a wider region

Fig. 2. Steady-state distribution of capital stock.
(reflecting depreciation rates), and depends on the characteristic of
the velocity of capital flows (provided by β ) within the region. For
12
the parameter range for which the poverty trap may occur, since
1−γ2
γ −γ
2 1
> 0, we note that relation (14) implies that the smaller the 11
c̄1
ratio c2
is, the easier it is for the poverty trap to occur. Furthermore,
10
the procedure followed in the proof provides detailed information
regarding the local behavior of capital stock near a point z0 with
9
zero savings. Finally, note that capital stock can be identically zero
inside a ball of center z0 and radius ϱ, i.e. well inside the region
8
k

where c1 vanishes, but capital may start accumulating (still inside

the region where c1 = 0) on account of spatial effects and capital
flow from nearby regions, since marginal productivity inside this
ball is high. Thus a poverty core might exist that is ‘‘surrounded’’ by
locations where savings vanish but due to capital flows, a positive
capital stock is accumulated at the steady state.
To provide a possible picture of the steady-state distribution
of the capital stock, we consider the solution of (13) under the
following parametrization for (9):
α = 0.4, δ = 0.03, ρ = 1.3, β = 0.7,
(15)
−z 2 / 4
D = 0.01, sA = 0.15e , z ∈ [−4, 4] .
The savings/productivity parameter sA reflects the assumption
that more developed locations in terms of per capita GDP have
a relatively higher savings/productivity parameter. As shown in
Fig. 2, the steady-state spatial distribution of the capital stock is
bell-shaped.
To obtain some insight into the emergence of poverty cores,
 withthe saving ratio defined as s (z ) =
we solve (13)
0.15 1 − exp −z 2 /4 , which suggests that savings vanish at the

center of the domain, and A = 1, which suggests no spatial
productivity differentiation. The emergence of the poverty core is
depicted in Fig. 3.
4.2. Time and space dependent solutions
Having studied the steady state, we now turn our attention to
the analysis of the full spatiotemporal Solow model. This means
finding the spatial distribution of capital stock k (t , z ) at each point
of time t that emerges from our spatial Solow growth model. The
corresponding mathematical problem then is in the general form
of Eq. (11).
Problem (11) presents an interesting technical twist, which is
directly related to economics. Since we are assuming decreasing
returns, α < 1, hence the function f is not Lipschitz continuous
for k, taking values in a neighborhood of zero, but rather Hölder
continuous. Even though this may sound like a boring technicality,
7
6
5
4
–4 –3 –2 –1 0 1 2 3 4
z
Fig. 3. Poverty core at the steady state.
it has important consequences for the uniqueness of solutions to
(11), since the uniqueness theorem for the solution of PDEs needs
the assumption of Lipschitz continuity of the nonlinearities. We
should note that this nonuniqueness problem is not a problem
only of the model we propose here; it is also true for the standard
temporal Solow model, as well as the PDE Solow version proposed
by Boucekkine and coworkers (see e.g. Boucekkine et al., 2009). The
problem arises from the Cobb–Douglas production function and
not from the transport term, whatever this may be. In the purely
temporal case, this phenomenon is almost never discussed since
we are usually interested in the region close to the nonzero steady
state k∗ ̸= 0, and in this region f is Lipschitz and no problem
arises. For the standard diffusion case in which ρ + α = 1, the
maximum principle guarantees that k > 0 so again there is no need
to pay too much attention to this problem. However, the case we
consider here, the emergence of poverty traps, dictates the need to
seriously consider the pathological region k = 0, since as shown in
Proposition 4, for a steady state we may have regions where k > 0
and regions where a poverty core emerges and k = 0.
We have the following existence result.
Proposition 5. Let β ≥ 0. If k0 ≥ 0, there exists a solution to
problem (11) in the weak sense. In particular there exist two weak
solutions of (11), k (the minimal solution) and k (the maximal
solution) such that any weak solution k of (11) satisfies 0 ≤ k ≤ k ≤
 A. Xepapadeas, A.N. Yannacopoulos / Journal of Mathematical Economics 67 (2016) 125–137 131

k ≤ M∗ . If k0 > 0, the solution to (11) is unique. The proposition

remains true also for Dirichlet or periodic boundary conditions.

Proof. For the proof, see Appendix A.4 in Appendix. 

The asymptotic in time behavior of the solutions of system (11)

is not a very easy problem because of: (a) the degeneracy of the
problem resulting from the inclusion of the nonlinear diffusion
term, and (b) the non-Lipschitz property of the nonlinearity which
is inherited by the use of the Cobb–Douglas production function.
Its full treatment requires abstract techniques from nonlinear
analysis, beyond the scope of the present article. However, here we
wish to present a preliminary result in this direction, related to the
problem of linearized stability of the steady-state solution, which
is important in its own right, and which highlights the various
subtleties of problem (11).
Let k∗ = k∗ (z ) be a steady-state solution of (11), that is, k∗
satisfies −1Φ (k∗ ) + f (z , k) = 0, and consider solutions of the Fig. 4. Spatiotemporal distribution of capital.
time dependent problem (11) with initial condition k(0, z ) =
k∗ (z ) + ϵ u(0, z ), where ϵ u(0, z ) is a small initial perturbation
around the steady state. This initial condition evolves according to
the evolution equation (11). It is intuitively clear that if u(t , z ) → 0
for all z ∈ U as t → ∞, then the solution k(t , z ) of the full
system (11) will have the property that k(t , z ) → k∗ (z ) for all
z ∈ U as t → ∞, hence all initial perturbation around the steady-
state solution k∗ will die out and be eliminated as an effect of the
dynamics and the steady state will be asymptotically stable. If, on
the other hand, |u(t , z )| → ∞ for all z ∈ U as t → ∞, then the
dynamics of the system will have as an effect the attenuation of
the initial small disturbance around the steady state, the state of
the system for any time will deviate from k∗ and the steady state
will be unstable.
Fig. 5. The time path of the Sobolev norm.
The problem of stability or instability of the steady state
is usually approximated by using the linearized version of the 4.3. Numerical simulations
evolution equation (11) around the steady state k∗ and looking at
the spectrum of an appropriate elliptic eigenvalue problem which Having established existence of solutions, steady states, and
depends on the particular form of k∗ . If the principal eigenvalue stability properties for the steady states, we turn now to some
is positive, then we have instability, whereas if it is negative, simulation results to determine the shape of the spatiotemporal
we have stability. However, this argument relies on the fact that distribution of capital emerging for the spatial Solow model
we may use the Taylor expansion to approximate Φ (k∗ (z ) + under plausible parameter choice. Our simulations numerically12
ϵ u(t , z )) and f (k∗ (z ) + ϵ u(t , z )) for small ϵ by a linear form in solve model (9), using the same parameter choice as in (15)
u. This is clearly inappropriate for (11) since f is non-Lipschitz of Section 4.1. Fig. 4 depicts the spatiotemporal evolution of
the stock of capital with initial condition k (0, z ) = e−z /4 +
2
(let alone continuously differentiable) in the neighborhood of zero.
Therefore, if k∗ approaches zero, or even worse if k∗ develops a 0.01 sin(50π z ) − 0.0183156, z ∈ [−4, 4]. This is a bell-shaped
poverty trap as shown in Proposition 4, the linearization argument distribution chosen for the purpose of approximating, through the
does not apply and the issue of treating the stability of the steady initial conditions, a distribution which could potentially emerge
state k∗ for small initial perturbations becomes more involved and if we consider economic – not geographical – space, with the
requires special attention. distance defined in terms of GDP per capita differences. The
boundary conditions are of zero flux type or ∂ k (t , −4) /∂ z =
The following proposition provides a stability result for small
∂ k (t , 4) /∂ z = 0.
perturbations of a steady state k∗ , which is also valid in the case
Fig. 5 depicts the evolution of the Sobolev norm defined as
where a poverty trap may occur.
 2
4
∂ k̂ (t , z )

Sb (t ) = dt ,
Proposition 6. Let β ≥ 0 and k∗ be a steady state. −4 ∂z

(i) If k∗ > 0, the steady state is linearly asymptotically stable. where k̂ (t , z ) is the solution of (9) as depicted in Fig. 4.
(ii) If k∗ ≥ 0, i.e., when the steady state develops a poverty trap, the The convergence of the Sobolev norm to a fixed number
core of the trap is persistent. means that the spatial gradients remain constant after a certain
point in time, implying that the system converges to spatially
The proposition also remains true for Dirichlet or periodic
boundary conditions.
12 Wolfram Mathematica was used for the numerical simulations. The PDEs were
Proof. For the proof, see Appendix A.5 in Appendix.  solved for t ∈ [0, 1000].
132 A. Xepapadeas, A.N. Yannacopoulos / Journal of Mathematical Economics 67 (2016) 125–137

Fig. 6. A flat spatiotemporal distribution of capital. Fig. 7. Spatiotemporal evolution of a poverty trap.

nonhomogeneous distribution of the stock of capital. Furthermore

the peak of the distribution in Fig. 4 converges for t > 200 to
a fixed positive number. Combining this with the convergence
of the Sobolev norm suggests that the growth model converges
in a spatiotemporal sense to a nonhomogeneous bell-shaped
capital stock distribution. This result is consistent with our theory
about the stability of spatially nonhomogeneous steady states
and the steadystate of Fig. 3. Since per capita output is given
α
by ŷ (t , z ) = k̂ (t , z ) , per capita output also converges to a
spatially nonhomogeneous distribution.
The bell-shaped pattern remains when we assume that ρ <
0, i.e., an increase in the stock of capital at a given location will
reduce the tendency of the capital stock to seek locations of higher
productivity, and β < 0. If there is no spatial variability of the
parameter sA, i.e. sA = 1, the spatial distribution becomes flat with
or without spatially differentiated initial conditions. This result is
consistent with the steady-state result obtained in Section 4.1, and Fig. 8. The spatiotemporal evolution for the trade balance model.
is shown in Fig. 6.
Our numerical results suggest therefore that under the
differentiation k (0, z ) becomes flatter but does not disappear
plausible zero flux boundary conditions, spatial variability of
completely. The Sobolev norm for this model converges to 0.11.
the saving/productivity parameter is important in generating a
The comparison of this value with the value of 17.9 for the
persistent spatially nonhomogeneous distribution for the stock of
model corresponding to Fig. 4, suggests that the strong spatial
capital in the spatial Solow model. AK models when α = 1 also
gradients and spatial heterogeneity are more persistent in models
result in flat spatial distribution.
with nonlinear diffusion than models with linear diffusion. This
The spatiotemporal evolution of a poverty trap, emerging from
observation could provide some insights into the mechanisms
spatially flat initial conditions, can also be
 shown if we assume as driving the spatiotemporal evolution of capital stock.
in Section 4.1 that s (z ) = 0.15 1 − exp −z 2 /4 and A = 1. This
The numerical simulations seem to support the theory devel-
is shown in Fig. 7. The Sobolev norm converges, suggesting that the
oped in the context of a spatial Solow model regarding the spa-
poverty core is persistent.
tiotemporal evolution of the capital stock and the existence of
A spatially nonhomogeneous bell-shaped pattern emerges and
steady states for a plausible set of parameter values regarding sav-
persists in time with Dirichlet boundary conditions k (−4, t ) =
ings rates depreciation and production elasticity. Furthermore they
k (4, t ) = k0 ≥ 0, circle boundary conditions k (−4, t ) =
seem to suggest that capital flows characterized by capital seeking
k (4, t ), and with time dependent boundary conditions k (−4, t ) =
locations of high returns and an endogenous flow velocity, result
k (4, t ) = γ t or k (−4, t ) = k (4, t ) = eγ t , which may reflect the
in a persistent spatially nonhomogeneous distribution of capital
assumption that locations with low capital stock at the beginning
and per capita output across locations if, as it is plausible, there
may grow fast.13
are savings/productivity differentials across locations. This result
Finally, when we parametrize for the trade balance model,
holds under various types of boundary conditions.
which corresponds to β = 1 and D = 1, with zero flux
boundary conditions, the result shown in Fig. 8 is that initial spatial
5. Conclusions and possible extensions

Seeking to explore mechanisms underlying the temporal

13 The bell-shaped patterns with Dirichlet boundary conditions emerge even with
evolution of the cross sectional distribution of per capita capital
no spatial variability in the sA parameter and in the AK model. This is anticipated
since, in general, Dirichlet boundary conditions especially of the hostile boundary
and output across space, we develop a spatial growth model where
k (−4, t ) = k (4, t ) ≈ 0 tend to ‘‘force’’ the formation of nonspatially homogeneous saving rates are exogenous. Capital movements across locations
patterns. This result is also consistent with the steady-state analysis in Section 4.1. are governed by a mechanism where capital moves toward
 A. Xepapadeas, A.N. Yannacopoulos / Journal of Mathematical Economics 67 (2016) 125–137 133

locations of relatively higher marginal productivity, with a velocity Definition 1. A function u ∈ W 1,2 (U ) is called a weak solution of
determined by the existing stock of capital. Considering that the (16) if
spatial domain corresponds to economic space, we developed  
a local model where the fundamental growth equation of the J (u) := ∇ u · ∇ψ dz + f (z , u)ψ dz = 0, ∀ψ ∈ W 1,2 (U ).
Solow model is augmented by a nonlinear diffusion term which U U
characterizes spatial movements.
If we restrict our choices to test functions ψ ≥ 0, and the above
We show that the augmented Solow equation has a solution and
(a) holds for a function u ∈ W 1,2 (U ) as the inequality J (u) ≤ 0, we
that steady states exist. Furthermore, under diminishing returns,
say that u is a sub-solution, (b) holds for a function ū ∈ W 1,2 (U )
the growth process could lead under plausible assumptions to
as the inequality J (ū) ≥ 0, we say that ū is a super-solution of (16)
a stable spatially non-homogeneous distribution for per capita
respectively.
capital and income in the long run. Insufficient savings may lead
to the emergence of poverty cores where capital stock is depleted Proof of Proposition 3. Using the Kirchhoff transformation u =
in some locations. Stability analysis indicates that a steady state k1+β = kρ+α we bring the system (13) to the equivalent form
with poverty cores is stable. This suggests that economies can
persistently remain in the poverty core while economies in other − 1u = c1 uγ1 − c2 uγ2 , (17)
locations will have a positive capital stock. In the spatial Solow with homogeneous Neumann boundary conditions where
model, zero capital stock in some locations is consistent with the
long-run stability of the entire spatial distribution of the stock α α 1 1
γ1 = = , and γ2 = = .
of capital. Numerical simulations confirm our theoretical results. 1+β ρ+α 1+β ρ+α
Our approach, by endogenizing the velocity of the capital flow,
provides a rich environment for studying growth processes in a For existence we can use the method of sub and supersolutions
spatiotemporal context. Furthermore, by linking capital flows with (see Definition 1). If we express Eq. (17) as J (u) := −1u −
differences in the marginal productivity of capital across locations c1 uγ1 + c2 uγ2 = 0 in U with ∂∂n u = 0 on ∂ U, and recall that a
and not with differences in the stock of capital across locations, our (weak) sub(super) solution u (ū) is a W 1,2 (U ) function such that
approach seems to provide a more plausible mechanism for capital J (u) ≤ 0 (J (ū) ≥ 0) resp. with the inequalities considered in
flows across locations. The emergence of spatial distributions a weak sense (see e.g. Schmitt, 2007 and references therein). If
where persistent poverty cores coexist with locations where the a pair of sup and supersolutions u and ū exist such that u ≤ ū
stock of capital is high, is a potentially interesting result suggesting with |c1 uγ1 (z ) − c2 uγ2 (z )| ≤ φ(z ) for every z ∈ U and u(z ) ∈
that the nonlinear diffusion approach could support outcomes [u(z ), ū(z )] with φ ∈ L2 (U ), then there exists a solution of (17) u
which could be in line with observed situations. The obvious such that u ≤ u ≤ ū (see e.g. Theorem 2.3 in Schmitt, 2007 for the
extension is to incorporate the transport mechanism for capital case of weak solutions).
developed here into a Ramsey-type optimal growth model. A candidate for ū is a constant function, ū =M. It is easily seen
that J (M ) = M γ1 −c1 + c2 M γ2 −γ1 ≥ M γ1 −c̄1 + c 2 M γ2 −γ1 ,
 
 1/(γ2 −γ1 )
Appendix. Proofs c̄1
so that choosing ū = M = c2
we guarantee

A.1. Proof of Proposition 1 that ū is a supersolution. On the other hand, we look for u
in the form u = ϵφ1 , where φ1 > 0 is the eigenfunction
Proof. Consider any V ⊂ U which consists only of interior points related to the dominant eigenvalue λ1 > 0 of the problem
of U. Assuming only the integrability of J, and since anything that −1φ = λφ , with homogeneous Neumann boundary conditions,
moves in or out of V must definitely pass through its boundary ∂ V , and ϵ > 0 is to be chosen appropriately. Since J (ϵφ1 ) =
γ −1 γ −1

we may express the net inflow (I) and outflow of capital (O) from ϵφ1 λ1 − ϵ γ1 −1 c1 φ1 1 + ϵ γ2 −1 c2 φ1 2 and the only negative
V in terms of a surface integral, term is the middle term, we see that for ϵ > 0 small enough
J (ϵφ1 ) ≤ 0 hence u is a subsolution. Therefore, there exists a
  
I−O= J · ndS = (vk) · ndS = − divJ (t , z )dz , weak solution u ∈ W 1,2 (U ) such that ϵφ1 ≤ u ≤ M. The proof
∂V ∂V V
then concludes using a standard bootstrapping argument (see e.g.
where we also used Gauss’ divergence theorem. This leads to a Ambrosetti and Álvarez, 2011 Chapter 7): Since u ∈ W 1,2 (U ) by
book-keeping equation for V of the form the Sobolev embedding theorem, u ∈ Lr0 (U ) (where r0 = ∞
∂ if d = 2 and r0 = d2d if d > 2). Then we rewrite (17) as
  
−2
k(t , z )dz = − divJ (t , z )dz + (s(z )A(z )f (k(t , z )) −1u = f (x) with f (x) = c1 (x)u(x)γ1 − c2 (x)u(x)γ2 and, by the
∂t
properties of u, we see that f ∈ Lr0 /γ2 (U ). Then the analogue of the
V V V
− δ k(t , z )) dz , Agmon–Douglis–Nirenberg estimates for the Neumann problem
whereby dividing by vol(V ) and passing to the limit as vol(V ) → −1u = f (see e.g. Agmon et al., 1959 or Lemma 5.2 in Wang,
0, we obtain (3), upon observing that Assumption 1 leads to 1991) guarantees that u ∈ W 2,p (U ) and a further application
(dropping explicit dependence on (t , z )) of the Sobolev embedding theorem implies that u ∈ C µ (Ū ) for
some µ ∈ (0, 1). Since c1 , c2 are Hölder functions, we have by
f ′ (k) 1
 
J (t , z ) = A Bψ(k)kf ′′ (k) ∇ k + ′′ ∇A .  the previous estimates that f is Hölder, and considering once more
f (k) A the Neumann problem −1u = f with a Hölder right hand side
using the extension of the Schauder theory for such boundary
conditions (see e.g. Theorem 6.26 in Gilbarg and Trudinger, 2015),
A.2. Proof of Proposition 3 we conclude the higher regularity of u.
For uniqueness we need the extension of the classic results of
For the reader’s convenience we recall the definition of weak Brezis and Oswald (1986), which are valid for the Dirichlet case,
solution for an elliptic equation of the form to the Neumann case. The uniqueness is guaranteed if the function
f (x,u)
−1u + f (z , u) = 0, in U , (16) u → ϕ(x, u) = u is strictly decreasing for every z ∈ U for u ∈
n · ∇ u = 0, on ∂ U . [0, ∞) (see e.g. Theorem 2 in Morales-Rodrigo and Suárez, 2006).
Since ϕ(x, u) = c1 uγ1 −1 −c2 uγ2 −1 we can calculate ϕ (x, u) = (γ1 −
′
134 A. Xepapadeas, A.N. Yannacopoulos / Journal of Mathematical Economics 67 (2016) 125–137
 
γ −1
1)c1 uγ1 −2 − (γ2 − 1)c2 uγ2 −2 = (γ1 − 1)uγ1 −2 c1 − γ2 −1 c2 uγ2 −γ1 , B2 \ B1 , 0 ≤ r − ϱ ≤ ϱ, we choose ν = 1−γ 2
, and for this choice
1 2
from which the claims follow.  we see that it is sufficient to choose the parameters as in (14).
Poverty traps cannot develop if β ≤ 0, since then
A.3. Proof of Proposition 4 0 = 1u + g (u)u ≤ −1u + Ku,

Proof. We will show that if γ1 < γ2 < 1, and c1 vanishes on a as any solution u satisfies 0 ≤ u ≤ M∗ , and since the function
subset of U (of positive measure), then any solution u of (17) (hence u → g (z , u) := −c1 uγ1 −1 + c2 uγ2 −1 is strictly increasing, it is
k) develops a poverty trap, i.e., u is a non-trivial solution of (17) bounded above by a positive constant K for every u ∈ [0, M ∗ ]. An
that vanishes on a region of positive measure. Our argument relies application of the strong maximum principle for −1u + Ku ≥ 0
heavily on Delgado and Suárez (2000), who studied a very similar leads to the result that u cannot vanish anywhere in the interior of
system with Dirichlet boundary condition.14 U, therefore a poverty core may not develop. 
Without loss of generality assume that c1 vanishes inside a ball
of radius 2ϱ centered at x0 = 0. Let Bi = B(0, iϱ), i = 1, 2 observe A.4. Proof of Proposition 5
that U = B1 ∪ (B2 \ B1 ) ∪ (U \ B2 ) =: RI ∪ RII ∪ RIII , and consider
the function Ψ (z ) = 01B1 (z ) + ψ(z )1B2 \B1 (z ) + 11U \B2 (z ), where We need to recall the following definition.
1A is the indicator function of a set A, and ψ is a function such
∂ψ Definition 2. Let k0 ∈ L∞ (U ).15 The function k ∈ L∞ (UT ), k ≥ 0,
that ∂ n = 0 on ∂ B1 , to be specified soon. If we show that any
is called a (weak) solution of (11) if
solution of (17) u ≥ 0 satisfies u ≤ M Ψ for M large enough or
equivalently that W + := max(u − M Ψ , 0) = 0, then clearly any
 
positive solution will develop a poverty core, located within B2 . k(t , z )ψ(t , z ) dz = k0 (z )ψ(0, z ) dz
 1/(γ2 −γ1 ) U U
, then W + (z ) = 0 for z ∈ RIII . ∂ψ
c̄1  t 
If M ≥ M∗ = c2
+ k(s, z ) (s, z ) + Φ (k(s, z ))1ψ(s, z )
If ∇ W (z ) = 0 a.e. (i.e. W is constant) then W (z ) = 0 for
+ + +
0 U ∂t
every z ∈ U and we are done. To show that it suffices to show

that I := U |∇ W + |2 dz ≤ 0. Observe that by the choice of M and + f (k(s, z ))ψ(s, z ) dz ds,
Ψ (and using Green’s theorem),
 for 0 ≤ t < T and every
I = ∇ W · ∇ W + dz 
∂ψ 

U
ψ ∈ J := ψ ∈ C (UT ) : ψ ≥ 0, = 0 and
∂ n (∂ U )T
 
= ∇ u · ∇ W + dz + M 1Ψ W + dz
∂ψ
B2 B2 \B1

  , 1ψ ∈ L2 (UT ) .
=− c2 uγ2 dz + M 1Ψ W + dz ∂t
B2 B2 \B1
 If we substitute the equality above with ≥, resp. ≤, then we obtain
≤ (−c2 uγ2 + M ∆Ψ ) W + dz =: J , the concept of supersolution, resp. subsolution of (11). A function
B2 \B1 which is at the same time a supersolution and a subsolution is a
solution.
where for the second line we used the weak form of (17) with
W + ∈ W 1,2 (U ) as a test function and the facts that c1 = 0 in B2 Proof of Proposition 5. The proof uses a regularization argument,
and c2 ≥ 0, u ≥ 0. We will show that J ≤ 0. Indeed, according to which we approximate the non-Lipschitz (with
 respect to the variable k) function f by a sequence of Lipschitz
J = (−c2 uγ2 + M 1ψ)(u − M ψ)1{u−M ψ≥0} dz , functions {fϵ } which converges in a monotone fashion to f as ϵ →
B2 \B1 0, and in particular fϵ ↑ f . This is possible for any continuous
function. For the present case one such possible regularization
recalling that Ψ = ψ on B2 \ B1 , and since (u − M ψ)1{u−M ψ≥0} ≥ 0,
scheme is the sequence of functions
we simply need to show that (−c2 uγ2 + M 1ψ)1{u−M ψ≥0} ≤ 0.
c1 (z )ϵ α−1 k − c2 (z )k, 0 ≤ k ≤ ϵ,

Noting that since c2 ≥ 0,
fϵ (z , k) =
γ2
(−c2 u + M 1ψ)1{u−M ψ≥0} ≤ −c 2 M ψ γ2 γ2
+ M 1ψ =: S (ψ), c1 (z )kα − c2 (z )k k ≥ ϵ,

we need to choose ψ as the solution of the inequality S (ψ) ≤ 0 which is easily seen to satisfy the required properties. We now
∂ψ consider the approximate problem
in B2 \ B1 with boundary conditions ∂ n = 0 on ∂ B1 and ψ = 1
on ∂ B2 . We look for spherically symmetric solutions of the form ∂w
ψ(r ) = ϱ−ν (r − ϱ)ν , for ν > 1 as yet unspecified, which upon = 1Φ (w) + fϵ (w), in (U )T ,
∂t
substitution leads to the condition ∂ Φ (w)
= 0, on (∂ U )T , (18)
S := −c 2 M γ2 −1 + ν(ν − 1)ϱ−σ (r − ϱ)σ −2 ∂n
+ (d − 1)νϱ−σ r −1 (r − ϱ)σ −1 ≤ 0, w(0, z ) = k0 (z ) + ϵ, z ∈ U ,
which on account of the Lipschitz property of fϵ has a unique
where σ = ν − γ2 ν . Clearly by the signs of the coefficients the
solution for every ϵ > 0, that we will denote by wϵ . Furthermore,
first term must dominate, and matching exponents noting that in
the comparison principle holds for (18), meaning that if w and w̄
are a sub and supersolution of (18) with w(0, z ) ≤ w̄(0, z ), then
w(t , z ) ≤ w̄(t , z ) for every t ∈ [0, T ] and the same result holds
14 In fact, it turns out that since the argument relies on local considerations, only
minor modifications are required, however, it is reproduced here as it allows us to
obtain concrete conditions on the parameters of the system for the poverty trap to
exist, which are of interest from the economic point of view. 15 Meaning that the initial condition is essentially bounded on U.
 A. Xepapadeas, A.N. Yannacopoulos / Journal of Mathematical Economics 67 (2016) 125–137 135

for solutions (see e.g. Ruan, 1998). Using the comparison principle we have that
for (18) repeatedly, we may conclude that {wϵ } is a nonincreasing 
∂ γ2

∂ γ2
sequence in ϵ which is bounded below by 0, hence the limit k := I1 := u ϕ= ϵ γ2 v ϕ
UT ∂ t (B1 )T ∂t
limϵ→0 wϵ ≥ 0 is well defined and is also a weak solution of the
γ2 −1 ∂v
original problem (11). We can now show that any solution k of + ϵγ2 u0 ϕ + O(ϵ),
(11) must satisfy the inequality k ≤ k , hence k is the maximal ∂t
(U \B1 )T
solution. To show that, consider the function kϵ := max(k, ϵ)
  
which can be shown to be a subsolution of (18) so that kϵ ≤ wϵ by
I2 := ∇ u · ∇ϕ = ∇ u0 · ∇ϕ + ϵ ∇v · ∇ϕ,
U U UT
the comparison principle for (18). Furthermore, since by definition T  T
γ
k ≤ kϵ , we see that k ≤ kϵ ≤ wϵ for every ϵ > 0 and passing to I3 := c1 uγ1 ϕ = c1 u01 ϕ
the limit as ϵ → 0 we obtain the required result k ≤ k . Again, by UT (U )T
a comparison principle, it is straightforward to see that k ≤ M∗ . γ −1
+ ϵγ1 c1 u01 vϕ + O(ϵ),
The minimal solution is constructed in terms of the solution (U \B2 )T
of the regularized problem (18) with the sole difference that the γ
where we used the fact that c1 = 0 on B2 (so that c1 u01 on B2 ), and
initial condition is now replaced by k0 (instead of k0 + ϵ ). This
is again a well-posed problem with a unique solution which for
  
γ
any ϵ > 0 will be denoted by vϵ . Repeated applications of the I4 := − c2 uγ2 ϕ = − ϵ γ2 c2 v γ2 ϕ − c2 u02 ϕ
(U )T (B1 )T (U )T
comparison principle for this problem lead to the conclusion that 
{vϵ } is a nondecreasing family in ϵ > 0 bounded above, hence the γ 2 −1
− ϵγ2 c2 u0 vϕ + O(ϵ).
limit k := limϵ→0 vϵ ≥ 0 is well-defined. It can be shown further (U \B1 )T
that for any weak solution k of (11), it holds that k ≤ k, hence k Collecting all the above in (19) we obtain that
is the minimal solution.
∂ γ2
 
Suppose now that k0 ≥ ϵ0 > 0. Choose ϵ < ϵ0 and consider
 
γ γ
(∇ u0 · ∇ − c1 u01 + c2 u02 )ϕ + ϵ γ2 v + c2 v γ2 ϕ
the solution of the regularized problem (18) for this choice of UT (B1 )T ∂t
ϵ and initial condition k0 . This problem has a unique solution
γ −1 ∂v
 
and by comparison results as well as by the construction of the +ϵ γ2 u02 ϕ+ ∇v · ∇ϕ
approximate family {fϵ }, it is clear that fϵ (k) = f (k) so that the (D\B1 )T ∂t UT
  
solution of the original problem coincides with the solution of the γ −1 γ2 −1
− γ1 c1 u01 vϕ + γ2 c2 u0 vϕ + O(ϵ) = 0.
regularized problem and uniqueness thus follows.  (U \B2 )T (U \B1 )T

Note the different orders of magnitude in the expansion, which

A.5. Proof of Proposition 6 now overcomes the technical difficulties of linearizing around the
poverty core. The zeroth order term in ϵ is identified as the weak
(i) The proof of (i) uses arguments very similar to the arguments form of the steady-state PDE and since u0 is the steady state
used in the proof of (ii) in the region U \ B2 and is omitted. solution this term vanishes. When we are within the poverty core,
(ii) Assume a poverty trap occurs and using the notation of the i.e., in (B1 )T , the next significant order of magnitude which is ϵ γ2
proof of Proposition 4, we assume that c1 vanishes in B2 and that (recall that γ2 < 1) is activated yielding that
the steady-state solution u0 vanishes in B1 , where B1 ⊂ B2 ⊂ U. ∂ γ2
We first apply the Kirchhoff transformation on (11). A standard v + c2 v γ2 = 0, in (B1 )T . (20)
linearization argument for the stability in terms of the expansion ∂t
u(t , z ) = u0 (z ) + ϵv(t , z ) cannot be used, as it would lead to the Finally, outside the poverty core, i.e., in (U \ B1 )T , the next order of
evolution equation magnitude ϵ 1 is activated, yielding

γ −1 ∂v γ −1 ∂v
  
γ −1 γ −1 γ −1
γ2 u02 = 1v + γ1 u01 c1 v − γ2 u02 c2 v, γ2 u02 ϕ+ ∇v · ∇ϕ − γ1 c1 u01 vϕ
∂t (D\B1 )T ∂t UT (U \B2 )T

for v , which is problematic as its validity relies on a Taylor + γ2 c2 u02
γ −1
vϕ = 0.
expansion of the nonlinearity, which in turn relies on the property (U \B1 )T
that the nonlinearity is C 1 which clearly fails in the core of the
As this is true for any test function ϕ , choosing first a test function
poverty trap where u0 is zero, a fact which is further displayed in
γ −1 γ −1 concentrated on (B2 \ B1 )T , where c1 = 0, and then a test function
the fact that the potentials u02 and u01 blow up to infinity in concentrated on (U \ B2 )T , we get that
the core of the poverty trap, hence the linearized equation is not
well-posed. γ −1 ∂v γ −1
γ2 u02 = 1v − γ2 c2 u02 v, z ∈ B2 \ B1 (21)
To overcome the above difficulties, we resort to the weak form ∂t
of (11) (after the application of Kirchhoff’s transformation), setting
and
u = u0 + ϵv and using a test function φ ∈ W 1,2 (U ),
γ −1 ∂v γ −1 γ −1
∂ γ2 γ2 u02 = 1v + γ1 c1 u01 v − γ2 c2 u02 v, z ∈ U \ B2 . (22)
  
γ1 γ2 ∂t
u ϕ+ ∇ u · ∇ϕ = (c1 u − c2 u )ϕ. (19)
UT ∂t UT UT
The evolution of a small perturbation of the poverty trap up to
We use the notation AT := [0, T ] × A for any set A, we break UT as O(ϵ) is governed by the evolution equations (20)–(22). Within the
UT = [0, T ]×(B1 ∪(B2 \B1 )∪(D\B2 )) =: (B1 )T ∪(B2 \B1 )T ∪(U \B2 )T , poverty core, in B1 , and since c2 > 0, Eq. (20) can be explicitly
and consider each of the integrals in the weak solution separately. estimated as
Using the fact that u0 = 0 on B1 and the differentiability of uγi , ∂ γ2
i = 1, 2 around u0 whenever z ̸∈ B1 , since u0 ̸= 0 in this region, v = −c2 v γ2 ≤ −c 2 v γ2
∂t
136 A. Xepapadeas, A.N. Yannacopoulos / Journal of Mathematical Economics 67 (2016) 125–137

hence then Re(Λ) > σ . If we manage to find a supersolution w̄ of the

c
− γ2 t above problem for some σ > 0, then our claim is valid. Let us
v(t , z ) ≤ v(0, z )e 2 , t > 0, z ∈ B1 , choose w̄ = u0 . Substituting into J, using the fact that u0 is a steady
which leads to the conclusion that limt →∞ v(t , z ) = 0 for any state, we get that
z ∈ B1 , a fact that guarantees the exponential stability of the γ γ −γ1
J (u0 ) = u01 [(1 − γ1 )c1 − {(1 − γ2 )c2 + σ γ2 }u02 ]
poverty core. The above argument can be modified if v(0, x) γ1 γ2 −γ1
is not continuous slightly, by taking the spatial average of (20) ≥ u0 [(1 − γ1 )c 1 − {(1 − γ2 )c̄2 + σ γ2 }u0 ],
over B1 , and obtaining the exponential estimate ∥v(t , ·)∥Lγ2 (B1 ) ≤   γ −γ 1
c c
− γ2 t which for c 1 ̸= 0 and since u0 ≤ c̄1
2 1
is positive if
∥v(0, ·)∥Lγ2 (B1 ) e 2 .
2
The situation is similar in B2 \ B1 , where we rearrange (21) as γ2 − γ1
0<σ < c̄2 ,
∂v 1 1−γ γ2
= u0 2 1v − c2 v, z ∈ B2 \ B1 , (23)
∂t γ2 so that a possible choice for σ is e.g. σ ∗ =
γ2 −γ1
c̄2 (or in fact
2 γ2
which is still singular, but only on the boundary ∂ B1 . This happens σ =
∗ γ2 −γ1
− ε , for every ε > 0). If c 1 = 0, as would happen for
c̄2
2 γ2
since by continuity, u0 (z ) = 0 for z ∈ ∂ B1 , so that the diffusion example on ∂ B2 , then a different choice for supersolution w can be
coefficient vanishes for z ∈ ∂ B1 , however it is nonzero elsewhere. used, for example w(z ) = 1 − e−ζ |z | for ζ > 0 and large enough.
Furthermore, by the estimates in the proof of Proposition 4 we
know that u0 ≤ C Ψ and in the relevant region (B2 \ B1 ) we have
References
that Ψ (z ) = ϱ−ν (|z | − ϱ)ν , with ν = 1−γ2
, so that we have the
2
1−γ
estimate 0 ≤ u0 2 ≤ C | |z | − ϱ|2 for an appropriate constant C . Acemoglu, Daron, 2009. Introduction to Modern Economic Growth. MIT press.
Agmon, Shmuel, Douglis, Avron, Nirenberg, Louis, 1959. Estimates near the
Eq. (23) is a linear equation. Looking for separable solutions of the boundary for solutions of elliptic partial differential equations satisfying
form v(t , z ) = eλt w(z ), upon substituting into the equation yields general boundary conditions. I. Comm. Pure Appl. Math. 12 (4), 623–727.
Ambrosetti, Antonio, Álvarez, David Arcoya, 2011. An Introduction to Nonlinear
γ −1
− 1w + γ2 u02 (λ + c2 )w = 0. (24) Functional Analysis and Elliptic Problems, Vol. 82. Birkhauser.
Baldwin, Richard, Forslid, Rikard, Martin, Philippe, Ottaviano, Gianmarco, Robert-
If λ < 0 then v(t , z ) = eλt w(z ) will tend to 0 as t → ∞,
Nicoud, Frédéric, 2003. Public Policies and Economic Geography. PUP,
Princeton.
hence the poverty trap will be stable. This is indeed the case, since Baldwin, Richard E., Martin, Philippe, 2004. Agglomeration and regional growth.
multiplying (24) by w , integrating over B2 \ B1 , and rearranging we Handb. Reg. Urban Econ. 4, 2671–2711.
Baldwin, Richard E., Martin, Philippe, Ottaviano, Gianmarco IP., 2001. Global
obtain income divergence, trade, and industrialization: the geography of growth take-
   offs. J. Econ. Growth 6 (1), 5–37.
γ −1 γ −1
λγ2 u0 2 w2 = − |∇w|2 − γ2 u0 2 c2 w 2 , Barro, Robert J., Martin, Xavier Sala-i, 2004. Economic Growth. Mit press,
B2 \B1 B2 \B1 B 2 \B 1 Cambridge, Massachusettes.
Boucekkine, Raouf, Camacho, Carmen, Fabbri, Giorgio, 2013a. On the optimal
which, since c2 > 0, leads to the result that λ < 0. control of some parabolic partial differential equations arising in economics.
AMSE, WP 2013-Nr 34.
Finally, for z ∈ U \ B2 , we rearrange (22) as Boucekkine, Raouf, Camacho, Carmen, Fabbri, Giorgio, 2013b. Spatial dynamics and
∂v 1 1−γ γ1 γ −γ convergence: The spatial ak model. J. Econom. Theory 148 (6), 2719–2736.
= u0 2 1v + c1 u01 2 v − c2 v, z ∈ U \ B2 , Boucekkine, Raouf, Camacho, Carmen, Zou, Benteng, 2009. Bridging the gap
∂t γ2 γ2 between growth theory and the new economic geography: The spatial ramsey
model. Macroecon. Dyn. 13 (1), 20.
the difference now being that for z ∈ U \ B2 , we are way out Bourguignon, François, Morrisson, Christian, 2002. Inequality among world
of the core of the poverty trap and therefore there exists K > 0 citizens: 1820–1992. Amer. Econ. Rev. 727–744.
Brezis, Haïm, Oswald, Luc, 1986. Remarks on sublinear elliptic equations. Nonlinear
such that u0 (z ) > K for all such z. This is now a standard linear Anal. TMA 10 (1), 55–64.
diffusion equation, with non-vanishing diffusion coefficient whose Brito, Paulo, 2004. The dynamics of growth and distribution in a spatially
heterogeneous world.
asymptotic behavior is determined by a weighted eigenvalue Brock, William, Xepapadeas, Anastasios, 2008. Diffusion-induced instability and
problem. Looking for separable solutions of (22) of the form pattern formation in infinite horizon recursive optimal control. J. Econom.
v(t , z ) = e−Λt w(z ),16 we see that Λ and w must be solutions of Dynam. Control 32 (9), 2745–2787.
Brock, William, Xepapadeas, Anastasios, 2010. Pattern formation, spatial external-
the weighted eigenvalue problem ities and regulation in coupled economic–ecological systems. J. Environ. Econ.
Manag. 59 (2), 149–164.
(−∆ + A(z ))w = ΛW (z )w, (25) Brock, William A., Xepapadeas, Anastasios, Yannacopoulos, Athanasios N., 2014a.
γ 2 −1 γ 1 −1 γ 2 −1 Optimal agglomerations in dynamic economics. J. Math. Econom. 53, 1–15.
where A = γ2 c2 u0 − γ1 c1 u0 and W = γ2 u0 is a Brock, William A., Xepapadeas, Anastasios, Yannacopoulos, Athanasios N., 2014b.
potential and weight function. If the principle eigenvalue of the Spatial externalities and agglomeration in a competitive industry. J. Econom.
Dynam. Control 42, 143–174.
weighted eigenvalue problem (25) is positive then λ < 0 (recall Camacho, Carmen, Zou, Benteng, 2004. The spatial solow model. Econ. Bull 18, 1–11.
that λ = −Λ) hence we have linearized stability. The existence of Cano-Casanova, Santiago, López-Gómez, Julián, 2002. Properties of the principal
a principle eigenvalue follows from the results of Cano-Casanova eigenvalues of a general class of non-classical mixed boundary value problems.
J. Differential Equations 178 (1), 123–211.
and López-Gómez (2002) (see Theorem 1.1, op cit). To check the Carlson, Dean A., Haurie, Alain, Leizarowitz, Arie, 1991. Infinite Horizon Optimal
positivity of Λ, we consider the following simple argument, based Control: Deterministic and Stochastic Systems. Springer-verlag.
on the method of sub and supersolutions. According to Protter and Conley, Timothy G., Ligon, Ethan, 2002. Economic distance and cross-country
spillovers. J. Econ. Growth 7 (2), 157–187.
Weinberger (2012) (Chapter 2, Section 8),17 if W > 0 and for some Delgado, Manuel, Suárez, Antonio, 2000. On the existence of dead cores for
σ ∈ R we may find a function w̄ such that degenerate lotkavolterra models. Proc. Roy. Soc. Edinburgh Sect. A 130 (04),
743–766.
J (w̄) := (−∆ + A(z ) − σ W )w̄ ≥ 0, Desmet, Klaus, Rossi-Hansberg, Esteban, 2009. Spatial growth and industry age. J.
Econom. Theory 144 (6), 2477–2502.
Desmet, Klaus, Rossi-Hansberg, Esteban, 2010. On spatial dynamics*. J. Reg. Sci. 50
(1), 43–63.
Fujita, Masahisa, Mori, Tomoya, 2005. Frontiers of the new economic geography*.
16 Here for reasons that will become obvious soon we use the reparametrization Pap. Reg. Sci. 84 (3), 377–405.
λ = −Λ. Gilbarg, David, Trudinger, Neil S., 2015. Elliptic Partial Differential Equations of
17 Note that in this reference the convention is to work if the operator ∆ rather Second Order. Springer.
Isard, Walter, Liossatos, Panagis, 1973. On optimal development over space and
than ∆ which explains the change in the signs.
time. J. Reg. Anal. Policy 3 (1).
 A. Xepapadeas, A.N. Yannacopoulos / Journal of Mathematical Economics 67 (2016) 125–137
Isard, Walter, Liossatos, Panagis, 1973. Transport investment and optimal
space–time development. Pap. Reg. Sci. 31 (1), 31–50.
Isard, W., Liossatos, P., 1978. Spatial Dynamics and Optimal Space-Time Develop-
ment. North-Holland, Amsterdam.
Lucas, Robert E., 1990. Why doesn’t capital flow from rich to poor countries?. Amer.
Econ. Rev. 80 (2), 92–96.
Lucas, Robert E., 2003. Macroeconomic priorities. Amer. Econ. Rev. 93 (1), 1–14.
Martin, Philippe, Ottaviano, Gianmarco IP., 2001. Growth and agglomeration.
Internat. Econom. Rev. 42 (4), 947–968.
Milanovic, Branko, 2009. Global inequality recalculated: the effect of new 2005
ppp estimates on global inequality. World Bank Policy Research Working Paper
Series. Vol.
Morales-Rodrigo, C., Suárez, Antonio, 2006. Uniqueness of solution for elliptic
problems with nonlinear boundary conditions. Comm. Appl. Nonlinear Anal.
13 (3), 69–78.
137
Prasad, Eswar, Rajan, Raghuram, Subramanian, Arvind, 2007. The paradox of capital.
Finance Dev. 44 (1).
Protter, Murray H., Weinberger, Hans F., 2012. Maximum Principles in Differential
Equations. Springer Science & Business Media.
Quah, Danny T., 1996. Twin peaks: growth and convergence in models of
distribution dynamics. Econ. J. 1045–1055.
Ray, Debraj, 1998. Development Economics. Princeton University Press.
Ruan, W.H., 1998. Monotone iterative method for degenerate nonlinear parabolic
equations. Nonlinear Anal. TMA 34 (1), 37–63.
Schmitt, Klaus, 2007. Revisiting the method of sub-and supersolutions for nonlinear
elliptic problems. Electron. J. Differential Equations 377.
Vázquez, Juan Luis, 2006. The Porous Medium Equation: Mathematical Theory:
Mathematical Theory. Oxford University Press.
Wang, Xu-Jia, 1991. Neumann problems of semilinear elliptic equations involving
critical sobolev exponents. J. Differential Equations 93 (2), 283–310.