Romer 1986

Increasing Returns and Long-Run Growth
Author(s): Paul M. Romer

Reviewed work(s):
Source: Journal of Political Economy, Vol. 94, No. 5 (Oct., 1986), pp. 1002-1037
Published by: The University of Chicago Press
Stable URL: http://www.jstor.org/stable/1833190 .
Accessed: 06/12/2012 20:31
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp
.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact [email protected].
The University of Chicago Press is collaborating with JSTOR to digitize, preserve and extend access to Journal
of Political Economy.
http://www.jstor.org
This content downloaded by the authorized user from 192.168.82.202 on Thu, 6 Dec 2012 20:31:53 PM
All use subject to JSTOR Terms and Conditions
Increasing Returns and Long-Run Growth
Paul M. Romer
Universityof Rochester
This paper presents a fully specified model of long-run growth in

which knowledge is assumed to be an input in production that has
increasing marginal productivity. It is essentially a competitive equi-
librium model with endogenous technological change. In contrast to
models based on diminishing returns, growth rates can be increasing
over time, the effects of small disturbances can be amplified by the
actions of private agents, and large countries may always grow faster
than small countries. Long-run evidence is offered in support of the
empirical relevance of these possibilities.
I. Introduction
Because of its simplicity, the aggregate growth model analyzed by
Ramsey (1928), Cass (1965), and Koopmans (1965) continues to form
the basis for much of the intuition economists have about long-run
growth. The rate of return on investment and the rate of growth of
per capita output are expected to be decreasing functions of the level
of the per capita capital stock. Over time, wage rates and capital-labor
ratios across different countries are expected to converge. Conse-
quently, initial conditions or current disturbances have no long-run
effect on the level of output and consumption. For example, an exog-
This paper is based on work from my dissertation (Romer 1983). An earlier version
of this paper circulated under the title "Externalities and Increasing Returns in Dy-
namic Competitive Analysis.' At various stages I have benefited from comments by
James J. Heckman, Charles M. Kahn, Robert G. King, Robert E. Lucas, Jr., Sergio
Rebelo, Sherwin Rosen, Jose A. Scheinkman (the chairman of my thesis committee),
and the referees. The usual disclaimer applies. I gratefully acknowledge the support of
NSF grant no. SES-8320007 during the completion of this work.
[oudsl of Poldclal Economy, 1986} vol. 94, no. 5J
? 1986 by T he University of Chicago. All rights reserved. 0022-3808/86/9405-0009$01.50
1002
INCREASING RETURNS 1003
enous reduction in the stock of capital in a given country will cause

prices for capital assets to increase and will therefore induce an offset-
ting increase in investment. In the absence of technological change,
per capita output should converge to a steady-state value with no per
capita growth. All these presumptions follow directly from the as-
sumption of diminishing returns to per capita capital in the produc-
tion of per capita output.
The model proposed here offers an alternative view of long-run
prospects for growth. In a fully specified competitive equilibrium, per
capita output can grow without bound, possibly at a rate that is mono-
tonically increasing over time. The rate of investment and the rate of
return on capital may increase rather than decrease with increases in
the capital stock. The level of per capita output in different countries
need not converge; growth may be persistently slower in less devel-
oped countries and may even fail to take place at all. These results do
not depend on any kind of exogenously specified technical change or
differences between countries. Preferences and the technology are
stationary and identical. Even the size of the population can be held
constant. What is crucial for all of these results is a departure from the
usual assumption of diminishing returns.
While exogenous technological change is ruled out, the model here
can be viewed as an equilibrium model of endogenous technological
change in which long-run growth is driven primarily by the accumula-
tion of knowledge by forward-looking, profit-maximizing agents.
This focus on knowledge as the basic form of capital suggests natural
changes in the formulation of the standard aggregate growth model.
In contrast to physical capital that can be produced one for one from
forgone output, new knowledge is assumed to be the product of a
research technology that exhibits diminishing returns. That is, given
the stock of knowledge at a point in time, doubling the inputs into
research will not double the amount of new knowledge produced. In
addition, investment in knowledge suggests a natural externality. The
creation of new knowledge by one firm is assumed to have a positive
external effect on the production possibilities of other firms because
knowledge cannot be perfectly patented or kept secret. Most impor-
tant, production of consumption goods as a function of the stock of
knowledge and other inputs exhibits increasing returns; more pre-
cisely, knowledge may have an increasing marginal product. In con-
trast to models in which capital exhibits diminishing marginal produc-
tivity, knowledge will grow without bound. Even if all other inputs are
held constant, it will not be optimal to stop at some steady state where
knowledge is constant and no new research is undertaken.
These three elements-externalities, increasing returns in the pro-
duction of output, and decreasing returns in the production of new
1004 JOURNAL OF POLITICAL ECONOMY
knowledge-combine to produce a well-specified competitive equilib-

rium model of growth. Despite the presence of increasing returns, a
competitive equilibrium with externalities will exist. This equilibrium
is not Pareto optimal, but it is the outcome of a well-behaved positive
model and is capable of explaining historical growth in the absence of
government intervention. The presence of the externalities is essen-
tial for the existence of an equilibrium. Diminishing returns in the
production of knowledge are required to ensure that consumption
and utility do not grow too fast. But the key feature in the reversal of
the standard results about growth is the assumption of increasing
rather than decreasing marginal productivity of the intangible capital
good knowledge.
The paper is organized as follows. Section II traces briefly the his-
tory of the idea that increasing returns are important to the explana-
tion of long-run growth and describes some of the conceptual
difficulties that impeded progress toward a formal model that relied
on increasing returns. Section III presents empirical evidence in sup-
port of the model proposed here. Section IV presents a stripped-
down, two-period version of the model that illustrates the tools that
are used to analyze an equilibrium with externalities and increasing
returns. Section V presents the analysis of the infinite-horizon, con-
tinuous-time version of the model, characterizing the social optimum
and the competitive equilibrium, both with and without optimal taxes.
The primary motivation for the choice of continuous time and the
restriction to a single state variable is the ease with which qualitative
results can be derived using the geometry of the phase plane. In
particular, once functional forms for production and preferences
have been specified, useful qualitative information about the dynam-
ics of the social optimum or the suboptimal competitive equilibrium
can be extracted using simple algebra. Section VI presents several
examples that illustrate the extent to which conventional presump-
tions about growth rates, asset prices, and cross-country comparisons
may be reversed in this kind of economy.
II. Historical Origins and Relation to Earlier Work

The idea that increasing returns are central to the explanation of
long-run growth is at least as old as Adam Smith's story of the pin
factory. With the introduction by Alfred Marshall of the distinction
between internal and external economies, it appeared that this expla-
nation could be given a consistent, competitive equilibrium interpre-
tation. The most prominent such attempt was made by Allyn Young
in his 1928 presidential address to the Economics and Statistics sec-
tion of the British Association for the Advancement of Science
(Young 1969), Subsequent economists (e.g., Hicks 1960; Kaldor

1981) have credited Young with a fundamental insight about growth,
but because of the verbal nature of his argument and the difficulty of
formulating explicit dynamic models, no formal model embodying
that insight was developed.
Because of the technical difficulties presented by dynamic models,
Marshall's concept of increasing returns that are external to a firm but
internal to an industry was most widely used in static models, espe-
cially in the field of international trade. In the 1920s the logical consis-
tency and relevance of these models began to be seriously challenged,
in particular by Frank Knight, who had been a student of Young's at
Cornell.' Subsequent work demonstrated that it is possible to con-
struct consistent, general equilibrium models with perfect competi-
tion, increasing returns, and externalities (see, e.g., Chipman 1970).
Yet Knight was at least partially correct in objecting that the concept
of increasing returns that are external to the firm was vacuous, an
"empty economic box" (Knight 1925). Following Smith, Marshall, and
Young, most authors justified the existence of increasing returns on
the basis of increasing specialization and the division of labor. It is
now clear that these changes in the organization of production cannot
be rigorously treated as technological externalities. Formally, in-
creased specialization opens new markets and introduces new goods.
All producers in the industry may benefit from the introduction of
these goods, but thev are goods, not technological externalities.2
Despite the objections raised by Knight, static models of increasing
returns with externalities have been widely used in international
trade. Typically, firm output is simply assumed to be increasing, or
unit cost decreasing, in aggregate industry output. See Helpman
(1984) for a recent survey. Renewed interest in dynamic models of
growth driven by increasing returns was sparked in the 1960s follow-
ing the publication of Arrow's (1962) paper on learning by doing. In
his model, the productivity of a given firm is assumed to be an increas-
ing function of cumulative aggregate investment for the industry.
Avoiding the issues of specialization and the division of labor, Arrow
argued that increasing returns arise because new knowledge is discov-
ered as investment and production take place. The increasing returns
were external to individual firms because such knowledge became
publicly known.
To formalize his model, Arrow had to face two problems that arise
l For an account of the development of Young's ideas and of his correspondence

with Knight. see Blitch (1983).
2For a treatment of increasing returns based on specialization, see Ethier (1982).
Although the model there is essentially static, it demonstrates how specialization can be
introduced in a differentiated products framework under imperfect competition.
1oo6 JOURNAL OF POLITICAL ECONOMY
in any optimizing model of growth in the presence of increasing

returns. The first, familiar from static models, concerns the existence
of a competitive equilibrium; as is now clear, if the increasing returns
are external to the firm, an equilibrium can exist. The second prob-
lem, unique to dynamic optimizing models, concerns the existence of
a social optimum and the finiteness of objective functions. In a stan-
dard optimizing growth model that maximizes a discounted sum or
integral over an infinite horizon, the presence of increasing returns
raises the possibility that feasible consumption paths may grow so fast
that the objective function is not finite. An optimum can fail to exist
even in the sense of an overtaking criterion. In the model of Arrow
and its elaborations by Levhari (1966a, 1966b) and Sheshinski (1967),
this difficulty is avoided by assuming that output as a function of
capital and labor exhibits increasing returns to scale but that the mar-
ginal product of capital is diminishing given a fixed supply of labor.
As a result, the rate of growth of output is limited by the rate of
growth of the labor force. Interpreted as an aggregate model of
growth (rather than as a model of a specific industry), this model leads
to the empirically questionable implication that the rate of growth of
per capita output is a monotonically increasing function of the rate of
growth of the population. Like conventional models with diminishing
returns, it predicts that the rate of growth in per capita consumption
must go to zero in an economy with zero population growth.
The model proposed here departs from both the Ramsey-Cass-
Koopmans model and the Arrow model by assuming that knowledge
is a capital good with an increasing marginal product. Production of
the consumption good is assumed to be globally convex, not concave,
as a function of stock of knowledge when all other inputs are held
constant. A finite-valued social optimum is guaranteed to exist be-
cause of diminishing returns in the research technology, which imply
the existence of a maximum, technologically feasible rate of growth
for knowledge. This is turn implies the existence of a maximum feasi-
ble rate of growth for per capita output. Over time, the rate of growth
of output may be monotonically increasing, but it cannot exceed this
upper bound.
Uzawa (1965) describes an optimizing growth model in which both
intangible human capital and physical capital can be produced. In
some respects, the human capital resembles knowledge as described
in this paper, but Uzawa's model does not possess any form of increas-
ing returns to scale. Instead, it considers a borderline case of constant
returns to scale with linear production of human capital. In this case,
unbounded growth is possible. Asymptotically, output and both types
of capital grow at the same constant rate. Other optimizing models
took the rate of technological change as exogenously given (e.g., Shell
1967b). Various descriptive models of growth with elements similar to

those used here were also proposed during the 1960s (e.g., Phelps
1966; von Wiezsacker 1966; Shell 1967a). Knowledge is accumulated
by devoting resources to research. Production of consumption goods
exhibits constant returns as a function of tangible inputs (e.g., physi-
cal capital and labor) and therefore exhibits increasing returns as a
function of tangible and intangible inputs. Privately produced knowl-
edge is in some cases assumed to be partially revealed to other agents
in the economy. Because the descriptive models do not use explicit
objective functions, questions of existence are generally avoided, and
a full welfare analysis is not possible. Moreover, these models tend to
be relatively restrictive, usually constructed so that the analysis could
be carried out in terms of steady states and constant growth rate
paths.
Continuous-time optimization problems with some form of increas-
ing returns are studied in papers by Weitzman (1970), Dixit, Mirrlees,
and Stern (1975), and Skiba (1978). Similar issues are considered for
discrete-time models in Majumdar and Mitra (1982, 1983) and De-
chert and Nishimura (1983). These papers differ from the model
here primarily because they are not concerned with the existence of a
competitive equilibrium. Moreover, in all these papers, the technical
approach used to prove the existence of an optimum is different from
that used here. They rely on either bounded instantaneous utility U(c)
or bounds on the degree of increasing returns in the problem; for
example, the production function f(k) is assumed to be such that f(k)/k
is bounded from above. The results here do not rely on either of these
kinds of restrictions; in fact, one of the most interesting examples
analyzed in Section VI violates both of these restrictions. Instead, the
approach used here relies on the assumptions made concerning the
research technology; the diminishing returns in research will limit
the rate of growth of the state variable. A general proof that restric-
tions on the rate of growth of the state variable are sufficient to prove
the existence of an optimum for a continuous-time maximization
problem with nonconvexities is given in Romer (1986).
Because an equilibrium for the model proposed here is a competi-
tive equilibrium with externalities, the analysis is formally similar to
that used in dynamic models with more conventional kinds of exter-
nalities (e.g., Brock 1977; Hochman and Hochman 1980). It also has a
close formal similarity to perfect-foresight Sidrauski models of money
demand and inflation (Brock 1975) and to symmetric Nash equilibria
for dynamic games (e.g., Hansen, Epple, and Roberds 1985). In each
case, an equilibrium is calculated not by solving a social planning
problem but rather by considering the maximization problem of an
individual agent who takes as given the path of some endogenously
determined aggregate variable. In the conventional analysis of exter-

nalities, the focus is generally on the social optimum and the set of
taxes necessary to support it as a competitive equilibrium. While this
question is addressed for this growth model, the discussion places
more stress on the characterization of the competitive equilibrium
without intervention since it is the most reasonable positive model of
observed historical growth. One of the main contributions of this
paper is to demonstrate how the analysis of this kind of suboptimal
equilibrium can proceed using familiar tools like a phase plane even
though the equations describing the equilibrium cannot be derived
from any stationary maximization problem.
III. Motivation and Evidence

Because theories of long-run growth assume away any variation in
output attributable to business cycles, it is difficult to judge the empir-
ical success of these theories. Even if one could resolve the theoretical
ambiguity about how to filter the cycles out of the data and to extract
the component that growth theory seeks to explain, the longest avail-
able time series do not have enough observations to allow precise
estimates of low-frequency components or long-run trends. When
data aggregated into decades rather than years are used, the pattern
of growth in the United States is quite variable and is apparently still
influenced by cyclical movements in output (see fig. 1). Cross-country
comparisons of growth rates are complicated by the difficulty of con-
trolling for political and social variables that appear to strongly in-
fluence the growth process. With these qualifications in mind, it is
useful to ask whether there is anything in the data that should cause
economists to choose a model with diminishing returns, falling rates
of growth, and convergence across countries rather than an alterna-
tive without these features.
Consider first the long-run trend in the growth rate of productivity
or per capita gross domestic product (GDP). One revealing way to
consider the long-run evidence is to distinguish at any point in time
between the country that is the "leader," that is, that has the highest
level of productivity, and all other countries. Growth for a country
that is not a leader will reflect at least in part the process of imitation
and transmission of existing knowledge, whereas the growth rate of
the leader gives some indication of growth at the frontier of knowl-
edge. Using GDP per man-hour as his measure of productivity, Mad-
dison (1982) identifies three countries that have been leaders since
1700, the Netherlands, the United Kingdom, and the United States.
Table 1 reports his estimates of the rate of growth of productivity in
each country during the interval when it was the leader. When the
TABLE 1
PRODUCTIVITY GROW H RATES FOR LEADING COUN TRIES
Annual Average Compound

Growth Rate of (GDP
Lead Countrv Interval per Man-Hour (M)
Netherlands 1700-1785 -.07

United Kingdom 17815-120 .5
United Kingdom 1820-90 1.4
United States 1890-1979 2.3
S~ORC-MaI ddsoln (1982).
productivity growth rate is measured over intervals several decades

long and compared over almost 3 centuries, the evidence clearly sug-
gests that it has been increasing, not decreasing. The rate of growth of
productivity increases monotonically from essentially zero growth in
eighteenth-century Netherlands to 2.3 percent per year since 1890 in
the United States.
Similar evidence is apparent from data for individual countries
over shorter horizons. Table 2 reports growth rates in per capita GDP
for the United States over five subperiods from 1800 to 1978. (The
raw data used here are from Maddison [1979].) These rates also sug-
gest a positive rather than a negative trend, but measuring growth
rates over 40-year intervals hides a substantial amount of year-to-year
or even decade-to-decade variation in the rate of growth. Figure 1
presents the average growth rate over the interval 1800-1839 (for
which no intervening data are available) and for the subsequent 14
decades. Identifying a long-run trend in rates measured over decades
is more problematical in this case, but it is straightforward to apply a
simple nonparametric test for trend.
Table 3 reports the results of this kind of test for trend in the per
capita rate of growth in GDP for several countries using raw data
TABLE 2
PER CAPITA GROWTH IN THE UNITED STATES
Average Annual Compound

Growth Rate of Real
Interval per Capita GDP (c)
1800-1840 .58
1840-80 1.44
1880- 920 1.78
1920-60 1.68
1960-78 2.47
SOURC.E.-Raw data are troin Maddison (1979).
30
-
2.
0.6 -,o
06 -
Co
F-ItJ 1.-Axerage annru 1 compote nd growth rate of per capita GDP in the U~nited
States for the interval 1800- 1839 and fo~r14 subsequent decades. Data are taken from
Madldiso~n(1979).
from Maddison (1979). The sample includes all countries for which
continuous observations on per capita GDP are available starting no
later than 1870. As for the data for the United States graphed in
figure 1, the growth rates used in the test for trend are measured over
decades where possible. The statistic -zrgives the sample estimate of
the probability that, for any two randomly chosen decades, the later
decade has a higher growth rate.
Despite the variability evident from figure 1, the test for trend for
the United States permits the rejection of the null hypothesis of a
nonpositive trend at conventional significance levels. This is true even
though growth over the 4 decades from 1800 to 1839 is treated as a
single observation. However, rejection of the null hypothesis depends
critically on the use of a sufficiently long data series. If we drop the
observation on growth between 1800 and 1839, the estimate of Xi
drops from .68 to .63 and the p-value increases from .03 to .1 f.3 If we
further restrict attention to the 11 decades from ar 1870 to 1978,
drops to .56 and the p-value increases to .29, 50 it is not surprising that
studies that focus on the period since 1870 tend to emphasize the
' The p-value gives the probability of observing a value of ir at least as large as the
reported value under the null hypothesis that the true probability is .5.
TABLE 3
A TEST FOR TREND IN PER CAPITA GDP GROWTH RATES
Date of
First Number of
Observation Observations Ir p-Value
United Kingdom 1700 20 .63 .06

France 1700 18 .69 .01
Denmark 1818 16 .70 .02
United States 1800 15 .68 .03
Germany 1850 13 .67 .06
Sweden 1861 12 .58 .25
Italy 1861 12 .76 .01
Australia 1861 12 .64 .11
Norway 1865 12 .81 .002
Japan 1870 11 .67 .07
Canada 1870 11 .64 .12
NOTE.-It is the sample estimate for each country of the probability that, for any two growth rates, the later one is
larger. The p-value is the probability of observing a value of it at least as large as the observed value under the null
hypothesis that the true probability is .5. Except in the early years when data are sparse, per capita rates of growth of
GDP were measured over successive decades. (Only two observations on growth rates are available for France prior
to 1820; for the United Kingdom, only two prior to 1800; for the United States, only one from 1800 to 1840.) For the
calculation of the p-value, see Kendall (1962). Data are from Maddison (1979).
constancy of growth rates in the United States. Rejection does not

appear to depend on the use of the rate of growth in per capita GDP
rather than the rate of growth of productivity. Reliable measures of
the work force prior to 1840 are not available, but using data from
Kuznets (1971) for the period 1840-1960 and from the 1984 Eco-
nomic Report of the President for 1960-80, one can construct a simi-
lar test for trend in the rate of growth of productivity over successive
decades. The results of this test, iT equal to .64 with a p-value of .10,
correspond closely to those noted above for growth in per capita GDP
over the similar interval, 1840-1978.
Over the entire sample of 11 countries, the estimated value for ar
ranges from .58 to .81, with a p-value that ranges from .25 to .002.
Five out of 11 of the p-values are less than .05, permitting rejection at
the 5 percent level in a one-sided test of the null hypothesis that there
is a nonpositive trend in the growth rate; eight out of 11 permit
rejection at the 10 percent level.
For less developed countries, no comparable long-run statistics on
per capita income are available. Reynolds (1983) gives an overview of
the pattern of development in such countries. Given the paucity of
precise data for less developed countries, he focuses on the "turning
point" at which a country first begins to exhibit a persistent upward
trend in per capita income. The timing of this transition and the pace
of subsequent growth are strongly influenced by the variations in the
world economy. A general pattern of historically unprecedented
growth for the world economy is evident starting in the last part of the
1800s and continuing to the present. This general pattern is inter-
rupted by a significant slowdown during the years between the two
world wars and by a remarkable surge from roughly 1950 to 1973.
Worldwide growth since 1973 has been slow only by comparison with
that surge and appears to have returned to the high rates that pre-
vailed in the period from the late 180Os to 1914.
Although all less developed countries are affected by the worldwide
economy, the effects are not uniform. For our purposes, the key
observation is that those countries with more extensive prior develop-
ment appear to benefit more from periods of rapid worldwide growth
and suffer less during any slowdown. That is, growth rates appear to
be increasing not only as a function of calendar time but also as a
function of the level of development. The observation that more de-
veloped countries appear to grow relatively faster extends to a com-
parison of industrialized versus less developed countries as well. In
the period from 1950 to 1980, when official estimates for GDP are
generally available, Reynolds reports that the median rate of growth
of per capita income for his sample of 41 less developed countries was
2.3 percent, "clearly below the median for the OECD countries for
the same period" (p. 975).
If it is true that growth rates are not negatively correlated with the
level of per capita output or capital, then there should be no tendency
for the dispersion in the (logarithm of the)4 level of per capita income
to decrease over time. There should be no tendency toward conver-
gence. This contradicts a widespread impression that convergence in
this sense has been evident, especially since the Second World War.
Streissler (1979) offers evidence about the source of this impression
and its robustness. For each year from 1950 to 1974, he measures the
variance across countries of the logarithm of the level of per capita
income. In a sample of ex post industrialized countries, those coun-
tries with a level of per capita income of at least $2,700 in 1974, clear
evidence of a decrease in the dispersion over time is apparent. In a
sample of ex ante industrialized countries, countries with a per capital
income of at least $350 in 1950, no evidence of a decrease in the
variance is apparent. The first sample differs from the second be-
cause it includes Japan and excludes Argentina, Chile, Ireland,
Puerto Rico, and Venezuela. As one would expect, truncating the
sample at the end biases the trend toward decreasing dispersion (and
' Examining the dispersion in the logarithm of the level of per capita income, not
dispersion in the level itself, is the correct way to test for convergence in the growth
rates. If the rate of growth were constant across countries that start from different
levels, the dispersion in the logarithm of the levels will stay constant, but dispersion in
the levels will increase.
at the beginning toward increasing dispersion). When a sample of all

possible countries is used, there is no evidence of a decrease in vari-
ance, but the interpretation of this result is complicated by the chang-
ing number of countries in the sample in each year due to data limita-
tions.
Baumol (1985) reports similar results. When countries are grouped
into industrialized, intermediate, centrally planned, and less devel-
oped economies, he argues that there is a tendency toward conver-
gence in the level of productivity within groups, even though there is
no tendency toward overall convergence. The tendency toward con-
vergence is clear only in his group of industrialized economies, which
corresponds closely to the sample of ex post industrialized countries
considered by Streissler. In any case, he finds no obvious pattern in
his entire sample of countries; if anything, there is a weak tendency
toward divergence.5
The other kind of evidence that bears directly on the assumption of
increasing returns in production comes from growth accounting ex-
ercises and the estimation of aggregate production functions. Econo-
mists believe that virtually all technical change is endogenous, the
outcome of deliberate actions taken by economic agents. If so and if
production exhibits constant returns to scale, one would expect to be
able to account for the rate of growth of output in terms of the rates
of growth of all inputs. The difficulty in implementing a direct test of
this assertion lies in correctly measuring all the inputs to production,
especially for intangible capital inputs such as knowledge. In a com-
prehensive attempt to account for the rates of growth in output in
terms of rates of growth of all inputs, including human and nonhu-
man, tangible and intangible stocks of capital, Kendrick (1976) con-
cluded that rates of growth of inputs are not sufficient to explain the
rate of growth of output in the 40-year interval 1929-69. For various
sectors and levels of aggregation, the rate of growth of output is 1.06-
1.30 times the appropriate aggregate measure of the rate of growth
for inputs. This kind of estimate is subject to substantial, unquantified
uncertainty and cannot be taken as decisive support for the presence
of increasing returns. But given the repeated failure of this kind of
growth accounting exercise, there is no basis in the data for excluding
the possibility that aggregate production functions are best described
as exhibiting increasing returns.
' Baumol (1985) argues that the convergence he observes among the industrialized
countries results from a transmission process for knowledge that takes place among the
industrialized countries but does not extend to centrally planned or less developed
countries. He would not agree that the apparent convergence is an artifact of an ex post
choice of the industrialized countries. Since he does not treat this issue directly, it is
diftcult to resolve it from his data. He does admit that his groupings are "somewhat
arbitrary."
IV. A Simple Two-Period Model

Even in the presence of increasing returns and externalities, calculat-
ing a social optimum is conceptually straightforward since it is equiva-
lent to solving a maximization problem. Standard mathematical re-
sults can be used to show that a maximum exists and to characterize
the solution by means of a set of necessary conditions. Despite the
presence of global increasing returns, the model here does have a
social optimum. The next section illustrates how it can be supported
as a competitive equilibrium using a natural set of taxes and subsidies.
This optimum is of theoretical and normative interest, but it cannot
be a serious candidate for describing the observed long-run behavior
of per capita output. To the extent that appropriate taxes and sub-
sidies have been used at all, they are a quite recent phenomenon.
The model here also has an equilibrium in the absence of any
governmental intervention. Much of the emphasis in what follows
focuses on how to characterize the qualitative features of this subop-
timal dynamic equilibrium. Although it is suboptimal, the competitive
equilibrium does satisfy a constrained optimality criterion that can be
used to simplify the analysis much as the study of the social optimiza-
tion problem simplifies the analysis in standard growth models.
The use of a constrained or restricted optimization problem is not a
new approach to the analysis of a suboptimal dynamic equilibrium.
For example, it has been widely used in the perfect-foresight models
of inflation. Nonetheless, it is useful to describe this method in some
detail because previous applications do not highlight the generality of
the approach and because the dynamic setting tends to obscure its
basic simplicity. Hence, I start by calculating a competitive equilib-
rium for a greatly simplified version of the growth model.
Specifically, consider a discrete-time model of growth with two pe-
riods. Let each of S identical consumers have a twice continuously
differentiable, strictly concave utility function U(Ci, C2), defined over
consumption of a single output good in periods 1 and 2. Let each
consumer be given an initial endowment of the output good in period
1. Suppose that production of consumption goods in period 2 is a
function of the state of knowledge, denoted by k, and a set of addi-
tional factors such as physical capital, labor, and so forth, denoted by
a vector x.6 To restrict attention to a choice problem that is essentially
6 For most of the subsequent discussion, k will be treated as a stock of disembodied

knowledge, i.e., knowledge in books. This is merely an expositional convenience and is
not essential. For example, if one wants to assume that all knowledge is embodied in
some kind of tangible capital such as conventional physical capital or human capital, k
can be reinterpreted throughout as a composite good made up of both knowledge and
the tangible capital good.
one-dimensional, assume that only the stock of knowledge can be

augmented; the factors represented by x are available in fixed supply.
To capture the basic idea that there is a trade-off between consump-
tion today and knowledge that can be used to produce more con-
sumption tomorrow, assume that there is a research technology that
produces knowledge from forgone consumption in period 1. Because
the economy here has only two periods, we need not be concerned
with the problem that arises in an infinite-horizon model when con-
sumption grows too fast and discounted utility goes to infinity. Thus
we do not need diminishing returns in research to limit the rate of
growth of knowledge, and we can choose a simple linear technology
with units such that one unit of forgone consumption produces one
unit of knowledge. A more realistic diminishing returns research
technology is described in the infinite-horizon model presented in the
next section.
Since newly produced private knowledge can be only partially kept
secret and cannot be patented, we can represent the technology of
firm i in terms of a twice continuously differentiable production func-
tion F that depends on the firm-specific inputs ki and xc and on the
aggregate level of knowledge in the economy. If N is the number of
firms, define this aggregate level of knowledge as K-- I ki.
The first major assumption on the production function F(ki, K, xi) is
that, for any fixed value of K, F is concave as a function of ki and xi.
Without this assumption, a competitive equilibrium will not exist in
general. Once concavity is granted, there is little loss of generality in
assuming that F is homogeneous of degree one as a function of ki and
xi when K is held constant; any concave function can be extended to
be homogeneous of degree one by adding an additional factor to the
vector x if necessary (Rockafellar 1970, p. 67). McKenzie (1959) re-
fers to this additional factor as an entrepreneurial factor. It can be
interpreted as an accounting device that transforms any profits into
factor payments.
By the homogeneity of F in ki and xi and by the assumption that F is
increasing in the aggregate stock of knowledge, K, it follows that F
exhibits increasing returns to scale. For any t > 1,
F(vki, SK, Ax) > F(Jki, K, Maxi)= 4vF(kiK, xi).
The second major assumption strengthens this considerably. It re-

quires that F exhibit global increasing marginal productivity of knowl-
edge from a social point of view. That is, for any fixed x, assume that
F(k, Nk, x), production per firm available to a dictator who can set
economywide values for k, is convex in k, not concave. This
strengthening of the assumption of increasing returns is what distin-
ioi6 JOURNAL OF POLITICAL ECONOMY
guishes the production function used here from the one used in the
models of Arrow, Levhari, and Sheshinski.
The equilibrium for the two-period model is a standard competitive
equilibrium with externalities. Each firm maximizes profits taking K,
the aggregate level of knowledge, as given. Consumers supply part of
their endowment of output goods and all the other factors x to firms
in period 1. With the proceeds, they purchase output goods in period
2. Consumers and firms maximize taking prices as given. As usual,
the assumption that agents treat prices and the aggregate level K as
given could be rationalized in a model with a continuum of agents.
Here, it is treated as the usual approximation for a large but finite
number of agents. Because of the externality, all firms could benefit
from a collusive agreement to invest more in research. Although this
agreement would be Pareto-improving in this model, it cannot be
supported for the same reasons that collusive agreements fail in mod-
els without externalities. Each firm would have an incentive to shirk,
not investing its share of output in research. Even if all existing firms
could be compelled to comply, for example, by an economywide mer-
ger, new entrants would still be able to free-ride and undermine the
equilibrium.
Because of the assumed homogeneity of F with respect to factors
that receive compensation, profits for firms will be zero and the scale
and number of firms will be indeterminate. Consequently, we can
simplify the notation by restricting attention to an equilibrium in
which the number of firms, N, equals the number of consumers, S.
Then per firm and per capita values coincide. Assuming that all firms
operate at the same level of output, we can omit firm-specific sub-
scripts.
Let x denote the per capita (and per firm) endowment of the fac-
tors that cannot be augmented; let e denote the per capita endowment
of the output good in period 1. To calculate an equilibrium, define a
family of restricted maximization problems indexed by K:
P(K): max U(cI, c2)
kEE[O,e]
subject to cl c e - ke
C2 < F(k, K, x),
x c x
Since U is strictly concave and F(k, K, x) is concave in k and x for each

value of K, P(K) will have a unique solution k for each value of K. (The
solution for x is trivially i.) In general, the implied values for cl, c2,
and k have no economic meaning. If K differs from Sk, then F(k, K, x)
is not a feasible level of per capita consumption in period 2. Equilib-
rium requires that the aggregate level of knowledge that is achieved
in the economy be consistent with the level that is assumed when firms
make production decisions. If we define a function r: RJR. R that
sends K into S times the value of k that achieves the maximum for the
problem P(K), this suggests fixed points of r as candidates for equilib-
ria.
To see that any fixed point K* of r can indeed be supported as a
competitive equilibrium, observe that P(K*) is a concave maximization
problem with solution k* = K*IS, cl -e - k*, and c = F(k*, Sk*, x)*
Since it is concave, standard necessary conditions for concave prob-
lems apply. Let Y denote a Lagrangian for P(K*) with multipliers pl,
P2e and w:
= U(CI, C2) + pi(i - - c1) + p2[F(k, K, X) - C2] + W(I - X).
When an interior solution is assumed, familiar arguments show that p3

-DjU(c, c2*)forj = 1, 2, that p, = p2DF(k*, Sk*, ), and that w-
p2D3F(k*, Sk*>X).7 As always, the shadow prices w and p1can be inter-
preted as equilibrium prices. To see this, consider first the maximiza-
tion problem of the firm: maxkp2F(k,SO, x) - pik - w X. x Since the
firm takes both prices and the aggregate level Sk* as given, a trivial
application of the sufficient conditions for a concave maximization
problem demonstrates that k* and i are optimal choices for the firm.
By the homogeneity of F with respect to its first and third arguments,
profits will be zero at these values. Consider next the problem of the
consumer. Income to the consumer will be the value of the endow-
ment, I = pie + w I = p2F(k*, Sk*, x) + p& - 4*).(The second
equality follows from the homogeneity of F in k and x.) When the
necessary conditions p1 = DjU(cl , c4) from the problem P(K*) are
used, it follows immediately that cr and e are solutions to the prob-
lem max U(CI, 2) subject to the budget constraint p1c, + p2c ? I. Note
that the marginal rate of substitution for consumers will equal the
private marginal rate of transformation perceived by firms, D1U(cl
?2)/D2U(C, I4) = DjF(k*, Sk*, x). Because of the externality, this dif-
fers from the true marginal rate of transformation for the economy,
D1F(k*, Sk*, ) + SD2F(k*, Sk*, i).
Arguments along these lines can be used quite generally to show
that a fixed point of a mapping like r defined by a family of concave
problems P(K) can be supported as a competitive equilibrium with
externalities. The necessary conditions from a version of the Kuhn-
Tucker theorem generate shadow prices associated with any solution
to P(K). The sufficient conditions for the problems of the consumer
and the firm can then be used to show that the quantities from the
7 Here, D denotes a derivative, Di the partial derivative with respect to the ith ar-
gument.
ioi8 JOURNAL OF POLITICAL ECONOMY
solution will be chosen in an equilibrium in which these prices are

taken as given. Conversely, an argument similar to the usual proof of
the Pareto optimality of competitive equilibrium can be used to show
that any competitive equilibrium with externalities for this kind of
economy will satisfy the restricted optimality condition implicit in the
problem P(K) (Romer 1983). That is, if K* is an equilibrium value of
aggregate knowledge, then K*/S will solve the problem P(K*). Thus
equilibria are equivalent to fixed points of the function F.
This allows an important simplification because it is straightforward
to characterize fixed points of F in terms of the underlying functions
U and F. Substituting the constraints from P(K) into the objective and
using the fact that x will be chosen to be x, define a new function V(k,
K) = UV - k, F(k, K, i)). Because of the increasing marginal produc-
tivity of knowledge, V is not a concave function; but for any fixed K, it
is concave in K. Then the optimal choice of k in any problem P(K) is
determined by the equation D1 V(k,K) = 0. Fixed points of F are then
given by substituting Sk for K and solving DIV(k, Sk) 0. Given
functional forms for U and F, this equation can immediately be writ-
ten in explicit form. The analysis can therefore exploit a three-way
equivalence between competitive equilibria with externalities, fixed
points of F, and solutions to an explicit equation DIV(k, Sk) = 0.
The key observation in this analysis is that equilibrium quantities
can be characterized as the solution to a concave maximization prob-
lem. Then prices can be generated from shadow prices or multipliers
for this problem. The complete statement of the problem must be
sought simultaneously with its solution because the statement involves
the equilibrium quantities. But since P(K) is a family of concave prob-
lems, solving simultaneously for the statement of the problem and for
its solution amounts to making a simple substitution in a first-order
condition.
V. Infinite-Horizon Growth
A. Descriptionof the Model
The analysis of the infinite-horizon growth model in continuous time
proceeds exactly as in the two-period example above. Individual firms
are assumed to have technologies that depend on a path K(t), t ? 0,
for aggregate knowledge. For an arbitrary path K, we can consider an
artificial planning problem PO(K) that maximizes the utility of a repre-
sentative consumer subject to the technology implied by the path K.
Assume that preferences over the single consumption good take the
usual additively separable, discounted form, f U(c(t))e - 8bdt with 8 >
0. The function U is defined over the positive real numbers and can
have U(O) equal to a finite number or to - x, for example, when U(c)
= ln(c). Following the notation from the last section, let F(k(t), K(t),
x(t)) denote the instantaneous rate of output for a firm as a function
of firm-specific knowledge at time t, economywide aggregate knowl-
edge at time t, and the level of all other inputs at t. As before, we will
assume that all agents take prices as given and that firms take the
aggregate path for knowledge as given.
Additional knowledge can be produced by forgoing current con-
sumption, but the trade-off is no longer assumed to be one-for-one.
By investing an amount I of forgone consumption in research, a firm
with a current stock of private knowledge k induces a rate of growth k
= G(I, k). The function G is assumed to be concave and homogeneous
of degree one; the accumulation equation can therefore be rewritten
in terms of proportional rates of growth, ilk / g(Ilk), with g(y) = G(y,
1). A crucial additional assumption is that g is bounded from above by
a constant ct. This imposes a strong form of diminishing returns in
research. Given the private stock of knowledge, the marginal product
of additional investment in research, Dg, falls so rapidly that g is
bounded. An inessential but natural assumption is that g is bounded
from below by the value g(O) = 0. Knowledge does not depreciate, so
zero research implies zero change in k; moreover, existing knowledge
cannot be converted back into consumption goods. As a normaliza-
tion to fix the units of knowledge, we can specify that Dg(O) = 1; one
unit of knowledge is the amount that would be produced by investing
one unit of consumption goods at an arbitrarily slow rate.
Assume as before that factors other than knowledge are in fixed
supply. This implies that physical capital, labor, and the size of the
population are held constant. If labor were the only other factor in
the model, exponential population growth could be allowed at the
cost of additional notation; but as was emphasized in the discussion of
previous models, a key distinguishing feature of this model is that
population growth is not necessary for unbounded growth in per
capita income. For simplicity it is left out. Allowing for accumulation
of physical capital would be of more interest, but the presence of two
state variables would preclude the simple geometric characterization
of the dynamics that is possible in the case of one state variable. If
knowledge and physical capital are assumed to be used in fixed pro-
portions in production, the variable k(t) can be interpreted as a com-
posite capital good. (This is essentially the approach used by Arrow
[1962] in the learning-by-doing model.) Given increasing marginal
productivity of knowledge, increasing marginal productivity of a
composite k would still be possible if the increasing marginal produc-
tivity of knowledge were sufficient to outweigh the decreasing mar-

ginal productivity associated with the physical capital.
Within the restrictions imposed by tractability and simplicity, the
assumptions on the technology attempt to capture important features
of actual technologies. As noted in Section II, estimated aggregate
production functions do appear to exhibit some form of increasing
returns to scale. Assuming that the increasing returns arise because of
increasing marginal productivity of knowledge accords with the plau-
sible conjecture that, even with fixed population and fixed physical
capital, knowledge will never reach a level where its marginal product
is so low that it is no longer worth the trouble it takes to do research. If
the marginal product of knowledge were truly diminishing, this
would imply that Newton, Darwin, and their contemporaries mined
the richest veins of ideas and that scientists now must sift through the
tailings and extract ideas from low-grade ore. That knowledge has an
important public good characteristic is generally recognized.8 That
the production of new knowledge exhibits some form of diminishing
marginal productivity at any point in time should not be controver-
sial. For example, even though it may be possible to develop the
knowledge needed to produce usable energy from nuclear fusion by
devoting less than 1 percent of annual gross national product (GNP)
to the research effort over a period of 20 years, it is likely that this
knowledge could not be produced by next year regardless of the size
of the current research effort.
B. Existence and Characterizationof a Social Optimum

Before using necessary conditions to characterize the solutions to
either the social optimization problem, denoted as PS., or any of the
artificial optimization problems P (K), I must verify that these prob-
lems have solutions. First I state the problems precisely. Let ko denote
the initial stock of knowledge per firm for the economy. As in the last
section, I will always work with the same number of firms and con-
sumers. Because the choice of x = x is trivial, I suppress this argu-
ment, writingf(k, K) = F(k, K, x). Also, let i(k) = f(k, Sk) = F(k, Sk, t)
denote the globally convex (per capita) production function that
would be faced by a social planner. In all problems that follow, the
constraint k(t) 2 0 for all t ? 0 and the initial condition k(O) = ko will
be understood:
8 See, e.g. Bernstein and Nadiri (1983) for estimates from the chemical industry sug-
gesting that spillover effects can be quite large.
PS,: max { U(c(t))e 8'dt
subject to k(t) ( (h(t) )
Px(K): max { U(c(t))e 8'dt
subject to k(t) K C
k(t) k(t)
Note that the only difference between these two problems lies in the
specification of the production function. In the first case, it is convex
and invariant over time. In the second, it is concave but depends on
time through its dependence on the path K(t). I can now state the
theorem that guarantees the existence of solutions to each of these
problems.
THEOREM 1. Assume that each of U, f, and g is a continuous real-
valued function defined on a subset of the real line. Assume that U
and g are concave. Suppose that i(k) = f(k, Sk) satisfies a bound 9;(k)
c p. + k' and that g(z) satisfies the bounds 0 ? g(x) c at for real
numbers p., p, and a. Then if tp is less than the discount factor 8, PSC
has a finite-valued solution, and Pcx(K)has a finite-valued solution for
any path K(t) such that K(t) e K(O)eo'.
The proof, given in an appendix available on request, amounts to a
check that the conditions of theorem I in Romer (1986) are satisfied.
Note that if (x is less than 8 the inequality otp< 8 allows for p > 1. Thus
the socially feasible production function i can be globally convex in k,
with a marginal social product and an average social product of
knowled ge that increase without bound.
The analysis of the social planning problem PS3, in terms of a cur-
rent-valued Hamiltonian and a phase plane follows along familiar
lines (see, e.g., Arrow 1967; Cass and Shell 1976a, 1976b). Define H(k,
X) = max, U(c) + X{kg([ (k) - c]lk)}. For simplicity, assume that the
functions U, f, and g are twice continuously differentiable. The first-
order necessary conditions for a path k(t) to be a maximum for PS,
are that there exists a path X(t) such that the system of first-order
differential equations k = D2H(k, K) and A = 8X - D IH(k, A) are
satisfied and that the paths satisfy two boundary conditions: the initial
condition on k and the transversality condition at infinity, lim,
X(t)k(t)e8- = Wt
X)Prosing the necessity of' the transversality condition for a maximization problem
that is not concave takes relatively sophisticated mathematical methods. Ekeland and
xx~
so~ ~ ~ ~ ~~~~s _k= O
k
FIG, 2.-Geometry of the phase plane for a typical social optimum. Arrows indicate
directions of trajectories in different sections of the plane. The rate of change of the
stock of knowledge, k, is zero everywhere on or below the locus denoted by k =*O; SO
denotes the socially optimal trajectory that stays everywhere between the lines X = 0
and k = 0.
Under the assumption that limbo DU(c) = oo, maximizing over c in

the definition of H(k, X) implies that DU(c) = XDg([9(k) - c]/k)
whenever the constraint k ? 0 is not binding; otherwise, c = i(k). This
gives c as a function of k and X. Substituting this expression in the
equations for k and A gives a system of first-order equations that
depends only on k and X.
Because of the restriction that k be nonnegative, the plane can be
divided into two regions defined by k = 0 and k - 0 (see fig. 2). In
a convenient abuse of the terminology, I will refer to the locus of
points dividing these two regions as the k = 0 locus. Along this locus,
both the conditions c = i;(k)and DU(c) = XDg([f(k) - c]lk) must hold.
Thus the k = 0 locus is defined by the equation DU(9(k)) = X. By the
concavity of U, it must be a nonincreasing curve in the k-X plane.
As usual, the equation A = 0 defines a simple locus in the plane.
When the derivative DIH(k, X) is evaluated along the k = 0 locus, the
equation for A there can be written A/X = 8- D(k). If D9 increases
without bound, there exists a value of k such that D9(k) > 8 for all k
Scheinkman (1983) prove the necessity of the transversality condition for nonconcave
discrete-time problems. In continuous time, a proof that requires a local Lipschitz
condition is given by Aubin and Clarke (1979).
larger than k, and for all such k, the A = 0 locus lies above the k = 0
locus. It may be either upward or downward sloping. If 3; were con-
cave and satisfied the usual Inada conditions, i = 0 would cross k = 0
from above and the resulting steady state would be stable in the usual
saddle-point sense. Here, K = 0 may cross k = 0 either from above or
from below. If D9i(k) is everywhere greater than 8, the A = 0 locus lies
everywhere above the k = 0 locus, and k can be taken to be zero.
(This is the case illustrated in fig. 2.) Starting from any initial value
greater than k, the optimal trajectory (K(t), k(t)), t ? 0, must remain
above the region where k = 0. Any trajectory that crosses into this
region can be shown to violate the transversality condition. Conse-
quently, k(t) grows without bound along the optimal trajectory.
This social optimum cannot be supported as a competitive equilib-
rium in the absence of government intervention. Any competitive
firm that takes K(t) as given and is faced with the social marginal
products as competitive prices will choose not to remain at the optimal
quantities even if it expects all other firms to do so. Each firm will face
a private marginal product of knowledge (measured in terms of cur-
rent output goods) equal to D1f; but the true shadow price of capital
will be Dlf + SD2f > Dlf. Given this difference, each firm would
choose to acquire less than the socially optimal amount of knowledge.
C. Existence and Characterizationof the

CompetitiveEquilibrium
Under a general set of conditions, this economy can be shown to have
a suboptimal equilibrium in the absence of any intervention. It is
completely analogous to the equilibrium for the two-period model. As
in that model, it is straightforward to show that there is a three-way
equivalence between competitive equilibria, fixed points of the map-
ping that sends a path K(t) into S times the solution to PR,(K),and
solutions to an equation of the form DI V(k, Sk) = 0.10 In the infinite-
horizon case, this equation consists of a system of differential equa-
tions, which can be represented in terms of a phase plane, and a set of
boundary conditions.
To derive these equations, consider the necessary conditions for the
concave problem P,(K). Define a Hamiltonian, denoted as H to distin-
guish it from the Hamiltonian H for the social planning problem PS.:
10 An explicit proof of this result is given in Romer (1983). The method of proof is
exactly as outlined in the two-period model. A generalized Kuhn-Tucker theorem is
used to derive the necessary conditions that yield shadow prices for the maximization
problems P(K). Suppose K* is a fixed point. If the consumer and the firm are faced
with the shadow prices associated with P4(K*), the sufficient conditions for their max-
imization problems are shown to be satisfied at the quantities that solve P4(K*).
H(k, K, K) =max U(c) + Xkg(,) cj
Then the necessary conditions for k(t) to be a solution to PT(K) are

that there exists a path X(t)such that k(t) = D2H(k(t), X(t),K(t)) and X(t)
- 8X(t) - DIH(k(t), X(t), K(t)) and such that the paths k(t) and X(t)
satisfy the boundary conditions k(O) = ko and lime, X(t)k(t)e't = 0.
Substituting Sk(t) for K(t) yields an autonomous system of differen-
tial equations, k(t) = DH(k(t), X(t), Sk(t)), A(t)- X(t) - D1H(k(t), X(t),
Sk(t)), that can be characterized using the phase plane. The two
boundary conditions must still hold. Any paths for k(t) and X(t) that
satisfy these equations and the boundary conditions will correspond
to a competitive equilibrium, and all competitive equilibria can be
characterized this way.
Before considering phase diagrams, I must show that a competitive
equilibrium exists for some class of models. Standard results concern-
ing the existence of solutions of differential equations can be used to
prove that the equations for X and k determine a unique trajectory
through any point (ki,K) in the phase plane. The difficulty arises in
showing that for any given value of ko there exists some value of Xo
such that the transversality condition at infinity is satisfied along the
trajectory through (k(, A0). As opposed to the case in which these
equations are generated by a concave maximization problem known
to have a solution, there is no assurance that such a X( exists.
The basic idea in the proof that such a X0 exists, and hence that a
competitive equilibrium exists, is illustrated in example I from the
next section. To state the general result, I need additional conditions
that characterize the asymptotic behavior of the functions f and g.
This is accomplished by means of an asymptotic exponent as defined
by Brock and Gale (1969). Given a function h(y), define the asymptotic
exponent e of h as e = limo logljh(y)l. Roughly speaking, h(y) behaves
asymptotically like the power function ye. Also, recall that x is the
maximal rate of growth for k implied by the research technology.
THEOREM 2. In addition to the assumptions of theorem 1, assume
that UCf, and g are twice continuously differentiable. Assume also that
A(k) = f(k, Sk) has an asymptotic exponent p such that p > I and op <
&. Finally, assume that Dg(x) has an asymptotic exponent strictly less
than - I. Let k be such that DIf(k, Sk) >8 for allk > k. Then if ko > K,
there exists a competitive equilibrium with externalities in which c(t)
and k(t) grow without bound.
The proof is given in Romer (1983, theorem 3). The assumption on
the asymptotic growth of 9; is self-explanatory. The assumption on
the asymptotic exponent of Dg is sufficient to ensure the boundedness
of g. The condition on D1 f will be satisfied in most cases in which 9(k)
- ik, Sk) is convex. Examples of functions satisfying these assump-

tions are given in the next section.
Once the conditions for the existence of a competitive equilibrium
have been established, the analysis reduces once again to the study of
the phase plane summarizing the information in the differential
equations. In many respects, this analysis is similar to that for the
social optimum for this economy. The phase plane can once again be
divided into regions where k = 0 and k > 0. Since by definition ;i(k) =
f(k, Sk), the equations for c as a function of k and X will be identical to
those in the social optimum: DU(c) =Dg([f(k, Sk) - c]Ik)if k > 0, c =
f, Sk) if k = 0. As a result, the boundary locus for the region k = 0
will also be identical with that from the social optimum. The only
difference arises in the equation for A. Although the equality H(k, X)
= H(k, K, Sk) does hold, the derivatives DIH(k, A) and DIH(k, K, Sk)
differ. In the first case, a term involving the expression D9(k) = Dlf(k,
Sk) + SD2f(k, Sk) will appear. In the second case, only the first part of
this expression, Diflk, Sk), appears. Therefore, D1H(k, A) is always
larger than D fH(k, K, Sk). Consequently, the A = 0 locus for the
competitive equilibrium must lie below that for the social optimum.
As was true of the social optimum, the K = 0 locus can be either
upward or downward sloping. If Dlf(k, Sk) > 8 for all k greater than
some value k, the K = 0 locus will lie above k = 0 for values of k to the
right of k. Then the qualitative analysis is the same as that presented
for the social optimum. Starting from an initial value ko > k, the only
candidate paths for equilibria are ones that stay above the k = 0
region; as before, paths that cross into this region will violate the
transversality condition. A trajectory lying everywhere in the region
where k > 0 can fail to have k(t) grow without bound only if the
trajectory asymptotically approaches a critical point where A and k are
both zero, but no such point exists to the right of k. Hence, all the
trajectories that are possible candidates for an equilibrium have paths
for k(t) that grow without bound. The existence result in theorem 2
shows that at least one such path satisfies the transversality condition
at infinity.
D. WelfareAnalysis of the CompetitiveEquilibrium

The welfare analysis of the competitive equilibrium is quite simple.
The intuition from simple static models with externalities or from the
two-period model presented in Section III carries over intact to the
dynamic model here. In the calculation of the marginal productivity
of knowledge, each firm recognizes the private return to knowledge,
Dlf(k, Sk), but neglects the effect due to the change in the aggregate
level, SD2f(k, Sk); an increase in k induces a positive external effect
D2f(k, Sk) on each of the S firms in the economy. Consequently, the

amount of consumption at any point in time is too high in the compet-
itive equilibrium and the amount of research is too low. Any interven-
tion that shifts the allocation of current goods away from consump-
tion and toward research will be welfare-improving. As in any model
with externalities, the government can achieve Pareto improvements
not available to private agents because its powers of coercion can be
used to overcome problems of shirking.
If the government has access to lump-sum taxation, any number of
subsidy schemes will support the social optimum. Along the paths
k*(t)and X*(t) from the social optimum, taxes and subsidies must be
chosen so that the first partial derivative of the Hamiltonian for the
competitive equilibrium with taxes equals the first partial derivative of
the Hamiltonian for the social planning problem; that is, the taxes
and subsidies must be chosen so that the after-tax private marginal
product of knowledge is equal to the social marginal product. This
can be accomplished by subsidizing holdings of k, subsidizing accumu-
lation k, or subsidizing output and taxing factors of production other
than k. The simplest scheme is for the government to pay a time-
varying subsidy of ri(t) units of consumption goods for each unit of
knowledge held by the firm. If this subsidy is chosen to be equal to the
term neglected by private agents, ul(t) = SD2f(k*(t), Sk*(t)), private
and social marginal products will be equal. A subsidy U2(t) paid to a
firm for each unit of goods invested in research would be easier to
implement but is harder to characterize. In general, solving for crAt)
requires the solution of a system of differential equations that de-
pends on the path for k*(t), In the special case in which production
takes the form f(k, K) = kVKY,the optimal subsidy can be shown to be
constant, O2 = -y/(v + y). (This calculation is also included in the app.
available on request.)
While it is clear that the social marginal product of knowledge is
greater than the private marginal product in the no-intervention
competitive equilibrium, this does not necessarily imply that interest
rates in the socially optimal competitive equilibrium with taxes will be
higher than in the suboptimal equilibrium. In each case, the real
interest rate on loans made in units of output goods can be written as
r(t) = - (p), where p(t) = e - btDU(c(t)) is the present value price for
consumption goods at date t. When utility takes the constant elasticity
form U(c) = [c - I/( - 0), this reduces to r(t) = 8 + O(Qc).In
the linear utility case in which 0 = 0, r will equal 8 regardless of the
path for consumption and in particular will be the same in the two
equilibria. This can occur even though the marginal productivity of
knowledge differs because the price of knowledge in terms of con-
sumption goods (equal to the marginal rate of transformation be-
tween knowledge and consumption goods) can vary. Holders of

knowledge earn capital gains and losses as well as a direct return equal
to the private marginal productivity of knowledge. In the case of
linear utility, these capital gains and losses adjust so that interest rates
stay the same.
This logical point notwithstanding, it is likely that interest rates will
be higher in the social optimum. On average, Uc will be higher in the
social optimum; higher initial rates of investment with lower initial
consumption must ultimately lead to higher levels of consumption. If
there is any curvature in the utility function U, so that 0 is positive,
interest rates in the optimum will be greater than in the no-
intervention equilibrium. In contrast to the usual presumption, cost-
benefit calculations in a suboptimal equilibrium should use a social
rate of discount that is higher than the market rate of interest.
VI. Examples
To illustrate the range of behavior possible in this kind of model, this
section examines specific functional forms for the utility function U,
the production function f, and the function g describing the research
technology. Because the goal is to reach qualitative conclusions with a
minimum of algebra, the choice of functional form will be guided
primarily by analytical convenience. For the production function, as-
sume thatf takes the form noted above,f(k, K) = kTK7. This is conve-
nient because it implies that the ratio of the private and social mar-
ginal products,
D f(k,_Sk) v
DI f(k, Sk) + SD2f(k, Sk) v + ly
is constant. Nonincreasing private marginal productivity implies that
o< v i 1; increasing social marginal productivity implies that 1 < -y
+ v. With these parameter values, this functional form is reasonable
only for large values of k. For small values of k, the private and social
marginal productivity of knowledge is implausibly small; at k = 0,
they are both zero. This causes no problem provided we take a mod-
erately large initial ko as given. An analysis starting from ko close to
zero would have to use a more complicated (and more reasonable)
functional form forf
Recall that the rate of increase of the stock of knowledge is written
in the homogeneous form k = G(I, k) = kg(JIk), where I is output
minus consumption. The requirements on the concave function g are
the normalization Dg(0) = I and the bound g(Ik) < for all Ilk. An
analytically simple form satisfying these requirements is g(z) = a(
+ z). Recalling that 8 is the discount rate, note that the bound re-
quired for the existence of a social optimum as given in theorem I

requires the additional restriction that o(v + y) < &. Given the stated
parameter restrictions, it is easy to verify that f and g satisfy all the
requirements of theorems I and 2.
A. Example I
With this specification of the technology for the economy, we can
readily examine the qualitative behavior of the model for logarithmic
utility U(c) =ln(c). The Hamiltonian can then be written as
H(k, X, K, c) = ln(c) + Xg (k, K)
Along (the boundary of the region in which) =4 0, Dg(O) = I implies

that c = A '.50 k = 0 is determined by the equation
-
X = [f(k, Sk)l - Skyk(v ).
The exact form for the locus =0 is algebraically complicated, but it
is straightforward to show that, for large k, X = 0 lies above the k = 0
locus since Dlf(k, Sk) will be greater than &. Also, if we define the
curve LI in the phase plane by the equation A = [1/(8 - a)]kk ' the A
= 0 locus must cross LI from above as indicated in figure 3. (Details
are given in the app. available on request.) Thus k - 0 behaves as k to
the power - (v + -y)< - 1, and A = 0 is eventually trapped between k
= 0 and a line described by k to the power - 1. In figure 3, represen-
tative trajectories t1 and t2 together with the competitive equilibrium
trajectory CE are used to indicate the direction of trajectories in the
various parts of the plane instead of the usual arrows.
-
Because the line LI is of the form K = [ 1/(8 - a)]k ' any trajectory
that eventually remains below LI will satisfy the transversality condi-
tion lim,.e tk(t)X(t) = 0. Given the geometry of the phase plane, it is
clear that there must exist a trajectory that always remains between
the loci A = 0 and k = 0. Given the initial value ko, index by the value
of Kall the trajectories that start at a point (ko, K) between the two loci.
The set of K'scorresponding to trajectories that cross = 0 can have
no smallest value, the set of K's that correspond to trajectories that
cross k = 0 can have no largest value, and the two sets must be
disjoint. Thus there exists a value K0such that the trajectory through
(ko, Kl) crosses neither locus and must therefore correspond to an
equilibrium."
This is the essence of the proof of theorem 2.
t2~~~~~~~~~~t
kO
0 1
o
k
Fiora 3.-Geometryof the competitive equilibrium for example 1. The line LI is
defined by the equation XA 1/(8 - o)k; t1 and t denote representative trajectories in
the phase plane, CE denotes the competitive equilibrium trajectory, which stays
everywhere between the A0 a(nd k = Oloci; Xo denotes the initial shadow price of
knowledge corresponding to the initial stock of knowledge ko.
In fact, the path resembles a conventional equilibrium in which the

trajectory remains between the A = 0 and k = 0 loci as it converges to
a saddle point, although here it is as if the saddle point has been
moved infinitely far to the right. Since the optimal trajectory cannot
stop, capital grows without bound. Since the trajectory is downward
sloping and since consumption is increasing in k and decreasing in X,
it is easy to see that consumption also grows without bound. Because
of the difficulty of the algebra, it is not easy to describe the asymptotic
rates of growth.
B. Example 2
Suppose now that utility is linear, U(c) = c. In the algebra and in the
phase plane for this case, we can ignore the restriction c 0 since it
will not be binding in the region of interest. Maximizing out c from
the Hamiltonian h(k, X, K, c) = c + Xkg((f - c)/k) implies that c = / -
3-
tk(X5 1). Thenf - c is positive (hence k is positive) if and only if
X> 1.
J ~~~~L2
I I*=o
_0 0~~~~~~~~~~~~~~
0 ko
k
FIG. 4.-Geometry of the competitive equilibrium for example 2. The line L2 is
defined by an equation of the form X = Ak"'Y-'; t, and t2 denote representative
trajectories in the phase plane; CE denotes the competitive equilibrium trajectory that
stays everywhere between L2 and A = 0; X0 denotes the initial shadow price of knowl-
edge.
In this example, it is possible to put tighter bounds on the behavior

of the A = 0 locus and, more important, on the behavior of the
equilibrium trajectory. As demonstrated in the appendix (available on
request). = 0 is upward sloping and behaves asymptotically like the
power function A = Bkv+" for some constant B. For this economy,
the equilibrium trajectory will lie above the A = 0 locus, so it is conve-
nient to define an additional curve that will trap the equilibrium tra-
jectory from above. For an appropriate choice of the constant A, the
line L2 defined by A = Akv+y- 1 will lie above A = 0 and will have the
property that trajectories must cross it from below (see fig. 4). Since
trajectories must cross A = 0 from above, the same geometric argu-
ment as used in the last example demonstrates that there exists a
trajectory that remains between these two lines. Consequently it must
also behave asymptotically like kv+Y- 1. Since k(t) can grow no faster
than eat, the product K(t)k(t) will be bounded along such a trajectory by
a function of the form eu(v+y)l. Since 8 > (v + y)x, this trajectory
satisfies the transversality condition and corresponds to an equilib-
rium.
Along the equilibrium trajectory, K behaves asymptotically like
0 I2
x~~~~~~~~~~~~~~~
X <
0~~~~~~~~~~~~~~~
Ao~~~~~~~~~~~~~~~~~~ /
0/
0~~~ ko
k
FIG. 5.-Geometry for the economy in example 2 when an exogenous increase of size
A in the stock of knowledge is known to occur at a time T > 0. The equilibrium
trajectory moves along t, until time T, at which point it is A units to the left of the
trajectory CE. At time T. the economy jumps horizontally to GE with the change in the
capital stock, but the path for A(t) is continuous. The equilibrium then proceeds along
CE. ko denotes the initial shadow price of knowledge in the case in which the exogenous
increase will take place; A0denotes the lower value that obtains in an economy in which
no exogenous increase will take place.
kV+- 1. Given the expression noted above for c in terms of A and k, c

behaves asvmptoticallv like kv+y - okl +(.5)(v+y- I) and I =f - be-
haves like k' + (5)(v+Y '1) Then c, I, Clk, and I/k go to infinity with k. By
the assumptions on the research technology, I/k going to infinity
implies that k/k approaches its upper bound a. Consequently, the
percentage rate of growth of output and of consumption will be in-
creasing, both approaching the asymptotic upper bound ot(v + -y).
Because the equilibrium trajectory is upward sloping, this economy
will exhibit different stability properties from either the conventional
model or the economy with logarithmic utility described above. Fig-
ure 5 illustrates a standard exercise in which a perfect-foresight equi-
librium is perturbed. Suppose that at time 0 it is known that the stock
of knowledge will undergo an exogenous increase of size A at time T
and that no other exogenous changes will occur. Usual arbitrage ar-
guments imply that the path for any price like X(t)must be continuous
at time T. The path followed by the equilibrium in the phase plane
starts on a trajectory like t, such that at time T it arrives at a point

exactly A units to the left of the trajectory CE from figure 4, which
would have been the equilibrium in the absence of any exogenous
change in k. As the economy evolves, it moves along tI then jumps A
units to the right to the trajectory CE at time T. Since e-btx(t) can be
interpreted as a time 0 market price for knowledge, a foreseen future
increase in the aggregate stock of knowledge causes a time 0 increase
in the price for knowledge and a consequent increase in the rate of
investment in knowledge. Because of the increasing returns, the pri-
vate response to an aggregate increase in the stock of knowledge will
be to reinforce its effects rather than to dampen them. Since the rate
of growth of the stock of knowledge is increasing in the level, this kind
of disturbance causes the stock of knowledge to be larger at all future
dates. Moreover, the magnitude of the difference will grow over time.
Thus small current or anticipated future disturbances can potentially
have large, permanent, aggregate effects.
As a comparison with the first example shows, this result requires
not only that increasing returns be present but also that marginal
utility not decrease too rapidly with the level of per capita consump-
tion. If we had restricted attention to the class of bounded, constant
elasticity utility functions, [cal-0) + 1]/(1 - 0) with 0 > 1, this phenom-
enon would not be apparent. The specific example here uses linear
utility for convenience, but similar results will hold for constant elas-
ticity utility function [c('-) - 1]/( 1 - 0) for values of 0 close enough
to zero.
C. Example 3
The analysis of the previous example suggests a simple multicountry
model with no tendency toward convergence in the level of per capita
output. Suppose each country is modeled as a separate closed econ-
omy of the type in example 2. Thus no trade in goods takes place
among the different countries, and knowledge in one country has
external effects only within that country. Even if all countries started
out with the same initial stock of knowledge, small disturbances could
create permanent differences in the level of per capita output. Since
the rate of growth of the stock of knowledge is increasing over time
toward an asymptotic upper bound, a smaller country s will always
grow less rapidly than a larger country 1. Asymptotically, the rates of
growth (k/k), and (k/k)1will both converge to o, but the ratios k/lk,and
c/ic will be monotonically increasing over time, and the differences
k1(t) - k,(t) and c,(t) - c,(t) will go to infinity.
It is possible to weaken the sharp separation assumed between
countries in this discussion. In particular, neither the absence of trade
in consumption goods and knowledge nor the sharp restriction on the

extent of the externalities is essential for the divergence noted above.
As in the Arrow (1962) learning-by-doing model, suppose that all
knowledge is embodied either in physical capital or as human capital.
Thus k denotes a composite good composed of both knowledge and
some kind of tangible capital. In this embodied form, knowledge can
be freely transported between two different countries. Suppose fur-
ther that the external effect of knowledge embodied in capital in
place in one country extends across its border but does so with dimin-
ished intensity. For example, suppose that output of a representative
firm in country 1 can be described asf(k, K1, K2) = kV(K' + Kb), where
k is the firm's stock of the composite good, K1 and K2 are the aggre-
gates in the two countries, and the exponent a on the domestic aggre-
gate K1 is strictly greater than the exponent b on the foreign aggregate
K2. Production in country 2 is defined symmetrically. Then for a
specific form of the research technology, Romer (1983) shows that the
key restriction on the equilibrium paths Sk1 and Sk2 in the two coun-
tries comes from the equality of the marginal product of private
knowledge imposed by the free mobility of the composite good k:
Dif(k1, Ski, Sk2) = Dlf(k2, Sk2, Skh). (1)
With the functional form given above, it is easy to verify that, in
addition to the symmetric solution kl = k2, there exists an asymmetric
solution. In that solution, if k1 is larger than k2 and growing (e.g.,
country 1 is industrialized and country 2 is not), the path for k2 that
satisfies this equation either can grow at a rate slower than that for
country 1 or may shrink, exporting the composite good to the more
developed country. 12
This kind of steady, ongoing "capital flight" or "brain drain" does
not require any fundamental difference between the two countries.
They have identical technologies. If we assume that there is perfect
mobility in the composite k, it can even take place when both countries
start from the same initial level of k. If all agents are convinced that
country 2 is destined to be the slow-growing country in an asymmetric
equilibrium, a discrete amount of the composite good will jump im-
mediately to country 1. Thereafter, the two countries will evolve ac-
cording to equation (1), with country 2 growing more slowly than
country 1 or possibly even shrinking.
This kind of model should not be taken too literally. A more real-
istic model would need to take account of other factors of production
with various degrees of less than perfect mobility. Nonetheless, it does
suggest that the presence of increasing returns and of multiple
12 Details are available in an app. available from the author.
equilibria can introduce a degree of instability that is not present in

conventional models. This identifies a second sense in which small
disturbances can have large effects. In addition to the multiplier-type
effect for a closed economy as described in the last example, a small
disturbance or a small change in a policy variable such as a tax rate
could conceivably have a decisive effect on which of several possible
equilibria is attained.
VII. Conclusion
Recent discussions of growth have tended not to emphasize the role
of increasing returns. At least in part, this reflects the absence of an
empirically relevant model with increasing returns that exhibits the
rigor and simplicity of the model developed by Ramsey, Cass, and
Koopmans. Early attempts at such a model were seriously under-
mined by the loose treatment of specialization as a form of increasing
returns with external effects. More recent attempts by Arrow,
Levhari, and Sheshinski were limited by their dependence on exoge-
nously specified population growth and by the implausible implica-
tion that the rate of growth of per capita income should be a mono-
tonically increasing function of the rate of population growth.
Incomplete models that took the rate of technological change as exog-
enously specified or that made it endogenous in a descriptive fashion
could address neither welfare implications nor positive implications
like the slowing of growth rates or the convergence of per capita
output.
The model developed here goes part way toward filling this theo-
retical gap. For analytical convenience, it is limited to a case that is the
polar opposite of the usual model with endogenous accumulation of
physical capital and no accumulation of knowledge. But once the
operation of the basic model is clear, it is straightforward to include
other state variables. The implications for a model with both increas-
ing marginal productivity of knowledge and decreasing marginal pro-
ductivity of physical capital can easily be derived using the framework
outlined here; however, the geometric analysis using the phase plane
is impossible with more than one state variable, and numerical
methods for solving dynamic equation systems must be used. 13 Since
the model here can be interpreted as the special case of the two-state-
variable model in which knowledge and capital are used in fixed
13
For an example of this kind of numerical analysis in a model with a stock of
knowledge and a stock of an exhaustible resource, see Romer and Sasaki (1985). As in
the growth model, increasing returns associated with knowledge can reverse conven-
tional presumptions; in particular, exhaustible resource prices can be monotonically
decreasing for all time.
proportions, this kind of extension can only increase the range of
possible equilibrium outcomes.
References
Arrow, Kenneth J. "The Economic Implications of Learning by Doing." Rev.
Econ. Studies 29 (June 1962): 155-73.
".Applications of Control Theory to Economic Growth." In Mathemat-
ics of0theDecision Sciences, vol. 1, edited by George B. Dantzig and Arthur F.
Veinott. Providence, R.I.: American Math. Soc., 1967.
Aubin, J. P., and Clarke, F. H. "Shadow Prices and Duality for a Class of
Optimal Control Problems." SIAMJ. Control and Optimization17 (Septem-
ber 1979): 567-86.
Baumol, William J. "Productivity Growth, Convergence and Welfare: What
the Long Run Data Show." Research Report no. 85-27. New York: New
York Univ., C. V. Starr Center, 1985.
Bernstein, J. I., and Nadiri, M. Ishaq. "Research and Development, Spillovers
and Adjustment Costs: An Application of Dynamic Duality at the Firm
Level." Working paper. New York: New York Univ., 1983.
Blitch, Charles P. "Allyn Young on Increasing Returns." J. Post Keynesian
Econ. 5 (Spring 1983): 359-72.
Brock, William A. "A Simple Perfect Foresight Monetary Model."J. Monetary
Econ. 1 (April 1975): 133-50.
. "A Polluted Golden Age." In Economicsof Natural and Environmental
Resources, edited by Vernon L. Smith. New York: Gordon and Breach,
1977.
Brock, William A., and Gale, David. "Optimal Growth under Factor Aug-
menting Progress."J. Econ. Theory 1 (October 1969): 229-43.
Cass, David. "Optimum Growth in an Aggregative Model of Capital Accumu-
lation." Rev. Econ. Studies 32 (July 1965): 233-40.
Cass, David, and Shell, Karl. "Introduction to Hamiltonian Dynamics in Eco-
nomics."J. Econ. Theory 12 (February 1976): 1-10. (a)
". The Structure and Stability of Competitive Dynamical Systems."J.
Econ. Theory 12 (February 1976): 31-70. (b)
Chipman, John S. "External Economies of Scale and Competitive Equilib-
rium." QJ.E. 84 (August 1970): 347-85.
Dechert, W. Davis, and Nishimura, Kazuo. "A Complete Characterization of
Optimal Growth Paths in an Aggregated Model with a Non-concave Pro-
duction Function." J. Econ. Theory31 (December 1983): 332-54.
Dixit, Avinash K.; Mirrlees, James A.; and Stern, Nicholas. "Optimum Saving
with Economies of Scale." Rev. Econ. Studies 42 (July 1975): 303-25.
Ekeland, Ivar, and Scheinkman, Jose A. "Transversality Conditions for Some
Infinite Horizon Discrete Time Maximization Problems." Technical Report
no. 411. Stanford, Calif.: Stanford Univ., IMSSS, July 1983.
Ethier, Wilfred J. "National and International Returns to Scale in the Model
Theory of International Trade." A.E.R. 72 (June 1982): 389-405.
Hansen, Lars Peter; Epple, Dennis; and Roberds, William. "Linear-Quadratic
Duopoly Models of Resource Depletion." In Energy, Foresightand Strategy,
edited by ThomasJ. Sargent. Washington: Resources for the Future, 1985.
Helpman, Elhanan. "Increasing Returns, Imperfect Markets, and Trade
Theory." In Handbookof InternationalEconomics,vol. 2, edited by Ronald W.
Jones and Peter B. Kenen. New York: North-Holland, 1984.
Hicks, John R. "Thoughts on the Theory of Capital: The Corfu Conference."

OxfordEcon. Papers 12 (June 1960): 123-32.
Hochman, Oded, and Hochman, Eithan. "Regeneration, Public Goods, and
Economic Growth." Econometrica48 (July 1980): 1233-50.
Kaldor, Nicholas. "The Role of Increasing Returns, Technical Progress and
Cumulative Causation in the Theory of International Trade and Economic
Growth." Econ. Apple.34, no. 4 (1981): 593-617.
Kendall, M. G. Rank CorrelationMethods. New York: Hafner, 1962.
Kendrick, John W. The Formation and Stocks of Total Capital. New York: Co-
lumbia Univ. Press (for N.B.E.R.), 1976.
Knight, Frank H. "On Decreasing Cost and Comparative Cost: A Rejoinder."
QJ.E. 39 (February 1925): 331-33.
Koopmans, Tjalling C. "On the Concept of Optimal Economic Growth." In
The Econometric Approach to Development Planning. Amsterdam: North-
Holland (for Pontificia Acad. Sci.), 1965.
Kuznets, Simon. "Notes on the Pattern of U.S. Economic Growth." In The
Reinterpretationof AmericanEconomicHistory, edited by Robert W. Fogel and
Stanley L. Engerman. New York: Harper and Row, 1971.
Levhari, David. "Extensions of Arrow's 'Learning by Doing.'" Rev. Econ.
Studies 33 (April 1966): 117-31. (a)
. "Further Implications of Learning by Doing." Rev. Econ. Studies 33
(January 1966): 31-38. (b)
McKenzie, Lionel W. "On the Existence of General Equilibrium for a Com-
petitive Market." Econometrica27 (January 1959): 54-71.
Maddison, Angus. "Per Capita Output in the Long Run." Kyklos32, nos. 1, 2
(1979): 412-29.
. Phases of CapitalistDevelopment.New York: Oxford Univ. Press, 1982.
Majumdar, Mukul, and Mitra, Tapan. "Intertemporal Allocation with a Non-
convex Technology: The Aggregative Framework." J. Econ. Theory 27
(June 1982): 101-36.
. "Dynamic Optimization with a Non-convex Technology: The Case of
a Linear Objective Function." Rev. Econ. Studies50 (January 1983): 143-5 1.
Phelps, Edmund S. "Models of Technical Progress and the Golden Rule of
Research." Rev. Econ. Studies 33 (April 1966): 133-45.
Ramsey, Frank P. "A Mathematical Theory of Saving." Econ.J. 38 (December
1928): 543-59.
Reynolds, Lloyd G. "The Spread of Economic Growth to the Third World:
1850-1980."J. Econ. Literature 21 (September 1983): 941-80.
Rockafellar, R. Tyrrell. Convex Analysis. Princeton, N.J.: Princeton Univ.
Press, 1970.
Romer, Paul M. "Dynamic Competitive Equilibria with Externalities, Increas-
ing Returns and Unbounded Growth." Ph.D. dissertation, Univ. Chicago,
1983.
"Cake Eating, Chattering and Jumps: Existence Results for Varia-
tional Problems." Econometrica,vol. 54 (July 1986).
Romer, Paul M., and Sasaki, Hiroo. "Monotonically Decreasing Natural Re-
source Prices under Perfect Foresight." Working Paper no. 19. Rochester,
N.Y.: Univ. Rochester, Center Econ. Res., 1985.
Shell, Karl. "A Model of Inventive'Activity and Capital Accumulation." In
Essays on the Theory of Optimal Growth, edited by Karl Shell. Cambridge,
Mass.: MIT Press, 1967. (a)
".Optimal Programs of Capital Accumulation for an Economy in
Which There Is Exogenous Technological Change." In Essayson the Theory
of Optimal Growth, edited by Karl Shell. Cambridge, Mass.: MIT Press,
1967. (b)
Sheshinski, Eytan. "Optimal Accumulation with Learning by Doing." In Es-
says on the Theoryof OptimalGrowth,edited by Karl Shell. Cambridge, Mass.:
MIT Press, 1967.
Skiba, A. K. "Optimal Growth with a Convex-Concave Production Function."
Econometrica46 (May 1978): 527-39.
Streissler, Erich. "Growth Models as Diffusion Processes: II. Empirical Illus-
trations." Kyklos32, no. 3 (1979): 571-86.
Weitzman, Martin L. "Optimal Growth with Scale Economies in the Creation
of Overhead Capital." Rev. Econ. Studies 37 (October 1970): 555-70.
von Weizsacker, Carl C. "Tentative Notes on a Two Sector Model with In-
duced Technical Progress." Rev. Econ. Studies 33 (July 1966): 245-51.
Young, Allyn A. "Increasing Returns and Economic Progress." In Readings in
Welfare Economics, edited by Kenneth J. Arrow and Tibor Scitovsky.
Homewood, Ill.: Irwin (for American Econ. Assoc.), 1969.
Uzawa, Hirofumi. "Optimum Technical Change in an Aggregative Model of
Economic Growth." Internat. Econ. Rev. 6 (January 1965): 18-31.

Romer 1986

Uploaded by

Copyright:

Available Formats

Romer 1986

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Romer 1986

Uploaded by

Copyright:

Available Formats

Increasing Returns and Long-Run Growth

Author(s): Paul M. Romer

This paper presents a fully specified model of long-run growth in

enous reduction in the stock of capital in a given country will cause

knowledge-combine to produce a well-specified competitive equilib-

II. Historical Origins and Relation to Earlier Work

(Young 1969), Subsequent economists (e.g., Hicks 1960; Kaldor

l For an account of the development of Young's ideas and of his correspondence

in any optimizing model of growth in the presence of increasing

1967b). Various descriptive models of growth with elements similar to

determined aggregate variable. In the conventional analysis of exter-

III. Motivation and Evidence

PRODUCTIVITY GROW H RATES FOR LEADING COUN TRIES

Annual Average Compound

Netherlands 1700-1785 -.07

S~ORC-MaI ddsoln (1982).

productivity growth rate is measured over intervals several decades

Average Annual Compound

SOURC.E.-Raw data are troin Maddison (1979).

A TEST FOR TREND IN PER CAPITA GDP GROWTH RATES

United Kingdom 1700 20 .63 .06

constancy of growth rates in the United States. Rejection does not

at the beginning toward increasing dispersion). When a sample of all

IV. A Simple Two-Period Model

6 For most of the subsequent discussion, k will be treated as a stock of disembodied

one-dimensional, assume that only the stock of knowledge can be

F(vki, SK, Ax) > F(Jki, K, Maxi)= 4vF(kiK, xi).

The second major assumption strengthens this considerably. It re-

Since U is strictly concave and F(k, K, x) is concave in k and x for each

When an interior solution is assumed, familiar arguments show that p3

solution will be chosen in an equilibrium in which these prices are

tivity of knowledge were sufficient to outweigh the decreasing mar-

B. Existence and Characterizationof a Social Optimum

gesting that spillover effects can be quite large.

PS,: max { U(c(t))e 8'dt

subject to k(t) ( (h(t) )

Px(K): max { U(c(t))e 8'dt

so~ ~ ~ ~ ~~~~s _k= O

Under the assumption that limbo DU(c) = oo, maximizing over c in

C. Existence and Characterizationof the

H(k, K, K) =max U(c) + Xkg(,) cj

Then the necessary conditions for k(t) to be a solution to PT(K) are

- ik, Sk) is convex. Examples of functions satisfying these assump-

D. WelfareAnalysis of the CompetitiveEquilibrium

D2f(k, Sk) on each of the S firms in the economy. Consequently, the

tween knowledge and consumption goods) can vary. Holders of

quired for the existence of a social optimum as given in theorem I

H(k, X, K, c) = ln(c) + Xg (k, K)

Along (the boundary of the region in which) =4 0, Dg(O) = I implies

This is the essence of the proof of theorem 2.

In fact, the path resembles a conventional equilibrium in which the

In this example, it is possible to put tighter bounds on the behavior

kV+- 1. Given the expression noted above for c in terms of A and k, c

starts on a trajectory like t, such that at time T it arrives at a point

in consumption goods and knowledge nor the sharp restriction on the

12 Details are available in an app. available from the author.

equilibria can introduce a degree of instability that is not present in

Hicks, John R. "Thoughts on the Theory of Capital: The Corfu Conference."