10dnote Devraj PDF

Notes for a Course in Development Economics
Debraj Ray
Version 3.35, 2010.

CHAPTER 1
Introduction
Open a book — any book —on the economics of developing countries, and it will begin with
the usual litany of woes. Developing countries, notwithstanding the enormous strides they
have made in the last few decades, display fundamental economic inadequacies in a wide
range of indicators. Levels of physical capital per person are small. Nutrition levels are low.
Other indicators of human capital such as education — both at the primary and seconday
levels — are well below developed-country benchmarks. So are access to sanitation, safe
water and housing. Population growth rates are high, and so are infant mortality rates. One
could expand this list indefinitely.
Notice that some of these indicators — infant mortality or life expectancy, for instance —
may be regarded as defining features of underdevelopment, so in this respect the list above
may be viewed, not as a statement of correlations, but as a definition of what we mean by
development (or the lack of it). But other indicators, such as low quantities of physical capital
per capita, or population growth rates, are at least one step removed. These features don’t
define underdevelopment. For instance, it is unclear whether low fertility rates are intrinsically
a feature of economic welfare or development. Surely, many families in rich countries may
take great pleasure in having a large number of offspring. Likewise, large holdings of
physical capital may well have an instrumental value to play in the development process,
but surely the mere existence of such holdings does not constitute a defining characteristic
of economic welfare.
And indeed, that is how it should be. We do not make a list of the features that go hand in
hand with underdevelopment simply to define the term. We do so because — implictly
or explicitly — we are looking for explanations. Why are underdeveloped countries
underdeveloped?1 It is easy enough to point to these inadequacies in terms of physical
and human capital, but the extra step to branding these as causes of underdevelopment is
perilously close, and we should avoid taking that step. Low levels of capital, or low levels
of education, are just as much symptoms of development as causes, and to the extent that
1
Perhaps the word “underdeveloped” does not constitute politically correct usuage, so that several publications
— those by well-known international organizations chief among them — use the somewhat more hopeful and
placatory present continuous “developing”. I won’t be using such niceties in this article, because it should be
clear — or at least it is clear in my mind — that economic underdevelopment pins no derogatory social label on
those who live in, or come from, such societies.
4 Introduction
they intertwine with and accompany the development process (or the lack of it), we cannot
rely on these observations as explanations.
That doesn’t stop economists from offering such explanations, however. More than one
influential study has regressed growth rates (alternatively, levels) of per-capita income on
variables such as the rate of savings and population growth. There is very little doubt, in fact,
that such variables are significantly associated with per-capita income. But nevertheless,
we do have to think about the sense in which these studies serve as explanations for
underdevelopment.
For instance, is it the case that individuals in different parts of the world have some intrinsic
difference in their willingness — or ability — to save, or to procreate? If this were the case,
we could hang our hat on the following sort of theory: such-and-such country is populated
by people who habitually save very little. This is why they are underdeveloped.
Somehow, this does not seem right. We would like to have a theory which — while not
belittling or downplaying the role of social, cultural and political factors — does not simply
stop there. We would like to know, for instance, whether low incomes provoke, in turn,
low savings rates so that we have a genuine chicken-and-egg problem. The same is true of
demographics — underdevelopment might be a cause of high population growth rates, just
as high population growth rates themselves retard the development process.
My goal in these notes is to talk about some of these chicken-and-egg situations, in which
underdevelopment is seen not as a failure of some fundamental economic parameters,
or socio-cultural values, but as an interacting “equilibrium” that hangs together, perhaps
precipitated by inertia or by history. [Indeed, in what follows, I will make a conceptual
distinction between equilibria created by inertia and those created by history.]
Why is this view of the development process an important one? There are three reasons
why I feel this view should be examined very seriously.
[1] This point of view leads to a theory, or a set of theories, in which economic “convergence”
(of incomes, wealth, levels of well-being) across countries is not to be automatically had.
Actually, the intelligent layperson reading these words will find this reasoning a bit abstruse:
why on earth would one expect convergence in the first place? And why, indeed, should I
find a theory interesting on the grounds that it does not predict convergence, when I knew
that all along? This is not a bad line of reasoning, but to appreciate why it is misguided, it is
important to refer to a venerable tradition in economics that has convergence as its very core
prediction. The idea is based — roughly — on the argument that countries which are poor
will have higher marginal products of capital, and consequently a higher rate of return to
capital. This means that a dollar of extra savings will have a higher payoff in poor countries,
allowing it grow faster. The prediction: pooere countries will tend to grow faster, so that
over time rich and poor countries will come together, or “converge”.
This is not the place to examine the convergence hypothesis in detail, as my intention is
to cover other views of development.2 But one should notice that convergence theories
in this raw form have rarely been found acceptable (though rarely does not mean never,
2
See Ray [1998], Chapters 2 and 3.
Introduction 5
among some economists), and there are several subtle variants of the theory. Some of
these variants still preserve the idea that lots of “other things” being equal, convergence
in some conditional sense is still to be had. It’s only if we start accepting the possibility
that — perhaps — these “other things” cannot be kept equal, that the notion of conditional
convergence starts losing its relevance and very different views of development, not at all
based on the idea of convergence, must be sought.
[2] The second reason why I find these theories important is that they do not reply on
“fundamental” differences across peoples or cultures. Thus we may worry about whether
Confucianism is better than the Protestant ethic in promoting hard-headed, succesful
economic agents, and we might certainly decry Hindu fatalism as deeply inimical to
purposeful, economic self-advancement, but we have seen again and again that when it
comes down to the economic crunch and circumstances are right, both Confucian and
Hindu will make the best of available opportunities — and so will the Catholics and a host
of other relgions and cultures besides. Once again, this is not the place to examine in detail
fundamentalist explanations based on cultural or religious differences, but I simply don’t
find them very convincing. This is not to say that culture — like conditional convergence
— does not play a role. [In fact, I provide such examples below.] But I also take the view
that culture, along with several other economic, social and political institutions, are all part
of some broader interactive theory in which “first cause” is to be found — if at all — in
historical accident.
[3] The last reason why I wish to focus on these theories is that create a very different role for
government policy. Specifically, I will argue that these theories place a much greater weight
on one-time, or temporary, interventions than theories that are based on fundamentals. For
instance, if it is truly Hindu fatalism that keeps Indian savings rates low, then a policy of
encouraging savings (say, through tax breaks) will certainly have an effect on growth rates.
But there is no telling when that policy can be taken away, or indeed, if it can be taken away
at all. For in the absence of the policy, the theory would tell us that savings would revert
to the old Hindu level. In contrast, a theory that is based on an interactive chicken-and-egg
approach would promote a policy that attempts to push the chicken-egg cycle into a new
equilibrium. Once that happens, the policy can be removed. This is not to say that once-and-
for-all policies are the correct ones, but only to appreciate that the interactive theories I am
going to talk about have very different implications from the traditional ones.
CHAPTER 2
Uneven Growth: A Research Agenda
The textbook paradigm of economy-wide development rests on the premise of “balanced

growth”; that is, on the presumption that all sectors will grow in unison over time as a
country gets richer. This view has served us reasonably well in several circumstances,
particularly those pertaining to macroeconomic models of long-term growth. An implicit
view that growth is balanced across sectors, or something close to it, also underlies the
notion of “trickle-down”, a stance that has strongly influenced development policy.
Of course, we would all agree that balanced growth is an abstraction. In many developing
countries, economic growth has been fundamentally uneven. First one sector, then another,
then a third have grown rapidly, but not all together. A list of some instances of this
phenomenon would include: software development, the outsourcing of services, quick
compositional shifts between agriculture and other sectors, the rise of export processing
zones, and others. The question really is not whether growth is balanced — it isn’t —
but whether the abstraction is a useful one. For many important development questions, I
believe the answer is no. This is why I would like to take the reality of “uneven growth”
seriously, and use it as an organizing device for a research program.
I divide my research agenda into roughly two parts: the sources and nature of uneven growth,
and the reactions to uneven growth. The first part studies the ways in which uneven growth
might arise, and its implications for economic inequality. The second part studies reactions:
how forces are set in motion to restore balancedness or perhaps even slow down or thwart
the growth process. To many, the former may appear unimportant without an appreciation
for the latter, so let me state at the outset that the second part is the more important section
of the paper, and the impatient reader is free to turn to it right away. But a few introductory
remarks may help as well.
In thinking about the effects of uneven growth, Albert Hirschman’s tunnel parable is useful
(see Hirschman and Rothschild (1973)). I present a slightly altered version. You’re in a
multilane tunnel, all lanes in the same direction, and you’re caught in a serious traffic jam.
After a while, the cars in the other lane begin to move. Do you feel better or worse? At first,
movement in the other lane may seem like a good sign: you hope that your turn to move will
come soon, and indeed that might happen. You might contemplate an orderly move into
the moving lane, looking for suitable gaps in the traffic. However, if the other lane keeps
whizzing by, with no gaps to enter and with no changes on your lane, your reactions may
8 Uneven Growth: A Research Agenda
well become quite negative. Unevenness without corresponding redistribution can be tolerated
or even welcomed if it raises expectations everywhere, but it will be tolerated for only so
long. Thus, uneven growth will set forces in motion to restore a greater degree of balance,
even (in some cases) actions that may thwart the growth process itself.
We could ignore this central issue. One reaction might be that we do not care about
distribution as long as the aggregates work right. Or perhaps some form of Coasian or
welfarist “compensation principle” is believed to be at work. Either reaction assumes away
or simply negates a crucial set of development problems, revolving around the political
economy of intersectoral or inter-group allocation.
The Hirschman parable also contains a parallel implication to which even less attention has
been paid. The movement of “neighboring lanes” under uneven growth not only brings
us information about what is possible, but it also defines and moulds our aspirations for
the future. Economists, mired as they generally are in a context-less description of human
preferences, are nowhere close to a theory of socially defined aspirations and for the double-
edged way in which they might influence individual behavior — either constructively, via
a profitable chain of investment and reward, or destructively, via frustration and violent
conflict.
Considerations such as these serve as entry points into some fundamental development
questions. Methodologically, they also underscore the need to look beyond traditional
models, by explicitly incorporating the social basis of individual preferences and well-being,
or by calling for better models of nonmarket allocations, such as those achieved through
lobbying and conflict.
As for the first part of this paper, on the sources of uneven growth, I have more precedent
to lean upon. Some of the earliest development models emphasized the varying roles of
different sectors of the economy. Rosenstein-Rodan’s (1943) view of underdevelopment
hinged on a failure of coordination across a variety of interlocking economic sectors.
Hirschman (1958) emphasized the concept of “leading sectors” that — by virtue of their
strong linkages to many other sectors — would pull the rest of the economy through the
development path. Nurkse (1953) and Lewis (1954) noted how agriculture might serve as a
near-inexhaustible supply of labor that might fuel industrial development without a drop
in per-capita food output. Rao (1952) and Ranis and Fei (1961) took these ideas further
by explicitly discussing a two-sector model that combined the working of a surplus-labor
agricultural sector with a demand-driven industrial sector. In some ways, my proposed
framework merely draws on a part of this earlier literature and marries it to modern
renditions of the aggregate growth model.
Our consideration of uneven growth leads to all sorts of questions that, in some ways, bridge
the gap between the micro and the macro of development. It also highlights a number of
ongoing issues in development economics in a unified way: notions of the dual economy
and the possibilities of trickle-down, theories of occupational choice, history-dependence,
the political economy of intersectoral allocation, socially determined aspirations, violent
conflict and the question of appropriate redistribution and compensation in the process of
development.
Uneven Growth: A Research Agenda 9
2.1 The Sources of Uneven Growth
Think of an economy populated by a large number of sectors, some final and some
intermediate. Each sector calls upon physical capital and various sorts of human capital to
produce output. As a starting point, imagine that all preferences and production change
smoothly and in the same proportions. Suppose further that there are no supply bottlenecks,
and that public resources are allocated evenly across all parts of the economy. In this
hypothetical economy, sectoral growth will simply track aggregate change in a balanced
way. In reality, these assumptions make little sense. Indeed, they miss out several problems
of fundamental interest. I begin by considering just why growth comes in uneven, sector-
specific jumps rather than in an even, well-diversified spray.
2.1.1 The Composition of Demand. The textbook assumption of microeconomic theory

is that an individual’s relative demand for different products is unaffected by her income.
This property is referred to as homotheticity. With homothetic demand, an increase in income
is always matched by a equiproportionate rise in the quantity demanded for every good.
Of course, demand isn’t homothetic in reality, and everyone knows it. As consumers, we
start off with the basics — food, clothing and shelter — and as our income grows, other
needs are met with other goods. As producers, we might shift to more methods that are
more intensive in technology or capital, as scale increases. The question is whether we can
get away with the homotheticity assumption as a simplifying device. Not surprisingly, the
answer depends on what we are interested in. For some macroeconomic questions, such as
the long-run rate of growth,1 the simplification may be a good one. For others, e.g., questions
of distribution across sectors or individuals, the assumption can be way off the mark.
Food is perhaps the most dramatic example of long-run uneven development, one that
necessitates basic structural change in the workforce, and in resource use.2 The majority
of India’s population lives and works in the rural sector. In the United States and in other
developed economies the corresponding percentage of the population is close to zero, and
in the corresponding Latin American percentages we see where India is surely headed
over the next decade or so. This transition is a fundamental source of unevenness which
informs much political and economic debate in many developing countries. While there are
attempts to integrate such nonhomotheticities into theories of distribution and trade,3 there
is to my knowledge little or no literature that ties these matters into the political economy
of development. I will return to this issue below.
To be sure, the general point goes beyond structural transformation from agriculture to
industry. The compositional pattern of demand — and its alteration as income grows — is
important whenever substantial costs must be incurred (including the cost of migration and
1
More generally, an assessment of the so-called “Kaldor facts” — the long-run constancy of growth rate, the
capital-output ratio, the real interest rate and so on — may not require much more than the one-sector growth
model.
2
Clark (1940) and Kuznets (1957) emphasized such structural change as fundamental aspects of the development
process. For models that address such aspects while attempting to retain conformity with the Kaldor facts, see,
e.g., Caselli and Coleman (2001), Kongsamut, Rebelo and Xie (2001) and Gollins, Parente and Rogerson (2007).
3
See, for instance, Markusen (1986), Baland and Ray (1991), Mani (2001), Matsuyama (2000, 2002), Foellmi and
Zweimuller (2006) and Fieler (2010).
relocation, not to mention training) to transfer factors of production one sector to another.
Agriculture and industry fit the bill, but so do, say, coal-mining and software.
The combination of nonhomothetic demand and costly transfer of human inputs can
reinforce or counteract each other. Perhaps the ultimate expression of reinforcement is
the dual economy. (The concept goes back at least to Nurkse (1953) and Lewis (1954).)
A dual economy is divided into near-independent enclaves. In the simplest version, a
subsistence sector produces basic consumption goods, and those employed in such sectors
consume these basics, with little left over for anything else. Side by side, a rich, sophisticated
subeconomy produces a variety of goods and services, largely for those employed in that
latter subeconomy. The two subeconomies coexist, but not necessarily in a symbiotic way.
If the high-tech subeconomy undergoes a boom, there are few connections for that boom to
transmit itself elsewhere. Trickle-down won’t work well in the dual economy.
Like all abstract concepts, the simplest version of the dual economy is a caricature. After
all, the rich do consume basics, and the poor are often employed in production of goods or
services consumed by those with higher incomes. But such intersectoral connections may
be relatively sparse. For that to be the case, the nonhomotheticity in demand must be of a
particular kind: by and large, the rich must consume things that the rich produce. Both the direct
and indirect demands (via the input-output matrix) of the rich must be intensive in factors
of production that the rich themselves supply — and a parallel observation must be true of
the poor.
Is this “segregation” something that we empirically observe? The question is of immense
importance in evaluating growth-based development processes that rely on trickle-down.
The dual economy relies on a particular sort of nonhomotheticity in direct and in derived
demand. We need empirically implementable metrics of the extent of “segregation”, that
measure the degree to which the dual-economy idea is true.
2.1.2 Uneven Growth and Inequality. The dual economy is a particularly dramatic
manifestation of unequal development. More generally, there are varying degrees of
segmentation rather than full segregation. This leads to the broader question of whether
and how uneven growth generates persistent inequalities, rather than trickle-down and
convergence.
There are several theories of the relationship between growth and inequality. For example,
in one approach based on aggregate growth models with diminishing returns to factors (in
the tradition of Solow (1957)), persistent inequality is entirely a matter of ongoing stochastic
shocks. Barring “luck”, these theories would predict no long-run inequality at all. Example
of this approach are Becker and Tomes (1979, 1986) and Loury (1981). A second approach
emphasizes presence of fixed costs in investment can generate steady state traps, in which
poverty breeds poverty (for example, Majumdar and Mitra (1982) and Galor and Zeira
(1993)). A third approach emphasizes occupational choice, a point to which I will return in
a later section.
The framework outlined in the previous section leads to a different, entirely complementary,
and equally significant view on persistent inequality. Growth leads to a change in the
commodity mix, which will then be translated into a change in the demand for factors.
Will this process be equalizing for incomes? Yes, if the original beneficiaries of the process
generate demands for factors held by individuals that did not originally benefit, setting in
motion a self-correcting change in the distribution of income and wealth. If, on the other
hand, the beneficiaries generate a demand mix that translate into demands for factors held
by those very same beneficiaries, the system must spiral away from equality. Our description
of the dual economy can now be viewed as a “steady state”, while the model of evolving
inequality sketched here represents an accompanying dynamic.
A proper empirical investigation of these conditions will require us to specify demand
systems that allow for compositional change (with income growth), and to estimate input-
output matrices that summarize different factor demands for different sectors.
Many of the central questions of development economics fall within this intuitive framework.
Issues of equitable and broad-based growth, the role of the market, the nature and scope of
government intervention: these can all be better studied with the basic model I have just
described.
2.1.3 Globalization and Inequality. The uneven growth framework is particularly useful
for thinking about globalization and its effects. It is a fairly general proposition that increased
openness heightens unevenness in production at the country level. After all, production in
sectors with a comparative advantage will grow faster as the economy opens to international
trade.
The effect of globalization on inequality is complex. There are low-income countries for
which an opening to trade will have textbook implications, such as the export of products
that are intensive in unskilled labor. There are middle-income developing countries with an
substantial supply of human capital for which the opposite is true: these will export goods
with a significant component of technical knowhow and skill. Both sets of countries will,
of course, exhibit factor-price equalization for factors active in the traded goods. But in the
former set, inequality will fall, while in the latter, inequality will rise. Indeed, if unskilled
labor is largely devoted to non-traded goods or services, the rise in inequality will be large
and noticeable. Globalization-induced uneven growth of this variety could well be a serious
concern.
A second layer of effects concerns cross-country inequality. Suppose that country-level
infrastructure is suitable for either high-tech or low-tech production, but not both. If both
high-tech and low-tech are important in world production and consumption, then some
country has to focus on low-tech and another on high-tech. Initial history will constrain
such choices, if for no reason than the fact that existing infrastructure (and national wealth)
determines the selection of future infrastructure. Over time, as the whole world climbs up
the income scale, the compositional change in demand will make for a greater proportion
of high-tech, and more and more countries will be able to make the transition. But on the
whole, if national infrastructure is more or less conducive to some (but not the full) range of
goods, there will be adverse distributional effects in the world economy as well.
This sort of model raises obvious questions. What is so specific about “national infrastruc-
ture”? Why is it not possible for every country to produces the same or similar mix of
goods, thus guaranteeing convergence? Do current national advantages somehow manifest
themselves in future advantages as well, thus ensuring that the world economy settles into
a permanent state of global inequality? As frameworks go, this is not a bad one to start
thinking about the effects of globalization.
As I have discussed in more detail in Ray (2007), it is possible to enrich this discussion
by bringing in the role of how early institutions set patterns for long-term growth (as in
Engerman and Sokoloff (1997), Sokoloff and Engerman (2000) and Acemoglu, Johnson and
Robinson (2001)). A specific colonial or institutional history will act as a constraint on —
or might mould — national infrastructure, driving the country into a particular production
slot in the global economy.
Our view of uneven growth fits in well with the entire debate on globalization. One
side of this debate emphasizes the convergence aspects: outsourcing, the establishment
of international production standards, technology transfer, political accountability, and the
spread of responsible macroeconomic policies. The persistent-inequality counterarguments
emphasize how a skewed playing field can only keep tipping. Nonconvexities and increasing
returns are endemic in this view. It is time to frame the globalization disputes within a
theoretical framework that allows for uneven growth, across sectors and across countries.
2.2 Reactions to Uneven Growth
If each sector were an island, uneven growth across them would not matter, at least to those
uninterested in inequality per se. But of course, there are many connections across sectors.
Organizing our thoughts along Hirschman’s parable, uneven growth initially provides hope,
in several ways. First, those who find themselves in a non-growth sector might be able to
participate in the sectors that are growing by changing the sort of human capital they have,
or — more realistically — changing the human capital their children will bring to the labor
market. This is the route of occupational choice. Second, the growing sectors might trickle
down — or over — and stimulate incomes in other sectors; this is the demand channel. Third,
government might use various supports, subsidies, salary adjustments, and protections,
to twist the river of economic progress through different territories. This is the avenue of
political economy, at least of the relatively peaceful kind. But there is also the possibility of
frustration: if uneven growth is perceived as persistent and exclusionary, society may enter
a different and darker realm: the possibility of development being scrapped or thwarted,
often by violent means; this is the area of conflict.
2.2.1 Occupational Choice. Consider information-technology-based services in a devel-

oping country such as Argentina or India: call centers, diagnostic services, legal services,
accounting and record-keeping, and so on. Initially, the process sets up a sizable demand
for those trained in the English language, as well as in computer literacy, basic business and
communication skills, and with some legal or medical training. The boom attracts attention
and raises sectoral incomes. How might this boom percolate to the rest of the country?
The first line of argument is occupational choice. In the context of our example, English-
medium education in India is going through a tremendous upsurge (see, e.g., Munshi
and Rosenzweig (2006)). Moreover, technical institutes are everywhere, vocational centers
are widespread and private business schools are commanding unprecedented prices. The
upsurge is even evident at primary and secondary schooling levels.
At some point, schooling will affect the composition of the work force, and in this way a
concentrated boom can spread. But there are limitations to what occupational choice can
achieve, and understanding these limitations and assessing their overall scope constitutes a
fertile research area.
First, changes in occupational structure can take a generation or more to implement. To
be sure, adults can sometimes retrain to match newly demanded skills, but this is typically
more the exception than the rule. This time lag means that there is substantial uncertainty in
sector-specific occupational choices. A student may train for an information-technology job,
but find that the economic pendulum has swung in a different direction — from software
and services to chipmaking or bioengineering, for instance — which requires a different
set of skills. In a rapidly changing world, the pursuit of generalized education may be
worthwhile, but of course generalized education has its limits in any particular field.
Second, and despite the time scales involved in occupational choice, it is often assumed
that parents will indirectly benefit, because they internalize the welfare of their children, or
perhaps because their children will look after them in old age. Such altruism indeed exists
(both ways), but it cannot serve as unqualified consolation for those whose boats will never
be lifted by the prosperity tide.
Finally, and perhaps the most-studied of these three points, there are the fundamental
difficulties of efficient occupational choice when capital markets are missing or imperfect
(for example, Banerjee and Newman (1993), Galor and Zeira (1993), Ljungqvist (1993),
Ghatak and Jiang (2002) and Mookherjee and Ray (2003)). This literature emphasizes the
history-dependence of the development process: an economy with high initial inequalities
could be trapped, unable to make the efficient sectoral choices and therefore losing out at the
aggregate level. While this literature does not emphasize uneven growth, it is easy enough
to see the beginnings of a line of research that combines capital-constrained occupational
choice models with sectoral growth.
2.2.2 Demand as a Channel for Spreading Income. A second avenue of percolation

is cross-sectoral demand. Suppose that there are only two kinds of agents in our
previous example: information technology service providers and potato growers, and that
information technology services is in the throes of a boom. How can the potato grower benefit
from the information technology boom? An obvious answer is that the service providers
will spend more on potatoes, leading to more income in the hands of an individual potato
grower.
We’ve already discussed some potential strengths and limitations of this effect (recall the
discussion on nonhomothetic preferences and trickle-down), but I now want to stress a
different point. Unless the supply of potatoes is completely elastic, the relative price of
potatoes must rise with the increase in demand. From a macroeconomic perspective — and
particularly when absolute prices in the boom sectors are sticky downwards — such relative
price “corrections” will generally cause some inflation. Ceteris paribus, an uneven-growth
economy must be more inflation-prone than a balanced-growth economy. In economies
where the boom takes place in the export sectors, this concern is multiplied, if there is interest
in keeping the domestic currency inexpensive to generate international competitiveness.
From a macroeconomic perspective, the connections between international currency man-
agement and domestic monetary policy are well-appreciated. From a perspective that
emphasizes uneven growth across sectors, there is more to learn. A particularly strict
stand on domestic inflation could hinder adequate relative price adjustments, and therefore
the percolation process. Inflation targeting in an unevenly growing economy is a delicate
business, and macroeconomists — with their focus on the aggregative model — have too
little theory with which to educate the central banker.
2.2.3 The Political Economy of Uneven Growth. Deliberate government policy can
generate comparative advantage in certain sectors, or at the very least it can help a boom
sector along once it starts. But which sectors are to be nurtured in this way, and which sectors
are to be abandoned, say, to existing foreign suppliers? Through most of the developing
world these questions are a subject of constant debate.
A key starting point is that government support cannot be provided to every activity at once.
Resources are limited, whether allocated to infrastructure, tax breaks, subsidies on the use
of electricity or fertilizer, or other forms. The allocation of public resources will ultimately
depend on debate, discourse and lobbying. At any one point of time, a small subset of
economic activities — perhaps “chosen” in an entirely random way — appear to be focal,
and the attention and energies of the country are drawn to it. Such is the case with outsourced
business in India from the United States and Europe: it accounts for a small fraction of Indian
employment, out of proportion to the attention and media interest it generates.
Yet the power to influence the media and the government depends on two things. The
first factor is the economic gain to be had from that influence. To the extent that sectoral
rankings in economic gain are correlated with their relative social values— a claim that will
not always hold true! — lobbying conveys useful information that can be used in efficient
resource allocation (for example, Austen-Smith (1994) and Austen-Smith and Banks (2002)).
The second factor is the wealth of potential lobbyists. While profitable sectors have more of
an incentive to lobby, sectors dominated by wealthy interest groups find it seasier to lobby.
Consequently, policymakers on the receiving end of such lobbies — even if they are honest
— can make bad resource-allocation decisions. Indeed, relative to this scenario, corruption
among policymakers may not make things that much worse (Esteban and Ray (2006)).
Indeed, uneven growth and the political economy of sectoral allocation together describe
the channel through which “reversals of fortune” in the sense of Acemoglu, Johnson and
Robinson (2002) might work. A country may have built up past wealth in certain sectors.
The interests in those sectors will become politically entrenched, and will resist change to
new sectors. The implicit argument here typically invokes corruption, but the argument is
broader than that: if information about newly profitable sectors is not common knowledge,
even a well-meaning government can succumb to entrenched interests, leading to a reversal
of fortune in high-inequality countries.
A large fraction of the everyday political economy of developing countries relates to one of
the most ubiquitous instances of uneven growth: specifically, the decline of agriculture as a
dominant livelihood for the majority, and the accompanying rise of industry and services.
The political economy issues here include agricultural subsidies, attitudes to globalization,
agrarian land rights, urban property rights, compensation related to dam placement, the
transfer of land from agrarian to commercial use, such as industry and cash crops. Much
recent research in development economics emphasizes these issues (for example, Besley
(1995), Binswanger, Deininger and Feder (1995), Stiglitz (2002), Goldstein and Udry (2008),
Duflo and Pande (2007), Field (2007) and Wang (2008)).
Two recent examples concern the proposed acquisition of land in Singur and Nandigram
(in West Bengal, India), the former to set up a car manufacturing plant for Tata, the latter
to set up a mega-chemical hub (one of hundreds of “Special Economic Zones” proposed all
around India). The proposed acquisitions amounted to less then 1 percent of overall arable
land in West Bengal. Yet political conflict over this move has been considerable. Intense
protests and agitations over the issue of land acquisition have resulted in violence and loss
of lives. Tata has since packed up and left, and the Nandigram plans have been shelved.
Other companies have sought greener pastures elsewhere in India.
The Singur-Nandigram examples raise issues that are relevant to structural transformation
in many developing countries. When trickle-down is minimal, or perceived to be so, the
question of how to compensate those who lose during economic transitions is of vital
importance. Property rights form a central aspect of this problem. As one instance, one
might think of well-defined, identifiable losers — say, a situation in which all land is farmed
by landowners and all factors of production are paid their opportunity cost. Then the owner
reaps all the residual surplus and his consent to sell should generally line up with social
efficiency. As another instance — typified by West Bengal — other agents have use rights to
the asset (sharecroppers in the case of land), and will also claim part of the surplus. In this
setting, official owners will be happy to sell land at a fraction of its true social value, because
much of that value has “leaked” to the individuals with use rights. It is inefficient, then, to
leave selling decisions to the owners alone.4
This argument can be extended to any activity in which non-owner agents receive a surplus
over and above the outside options for that agent. Efficient shutdown of that activity will
need the consent of those agents.
Next, the form of compensation deserves consideration. An excessive emphasis on one-off
compensation exposes recipients to a risk that unexpected high inflation may erode their
payments. Alternative proposals include compensation in the form of an inflation-adjusted
monthly pension combined with a savings bond that can be sold (so that people have access
to the equity value if a distress sale is required), payments linked to the future value of
transferred land, or even shares in the new companies.5
The final point concerns credibility; in particular, the setting-up of a structure to ensure that
promised payments are actually made. The history of credible compensation in India is
problematic, to say the least. There is room here for serious, applied exercises in mechanism
4
Ghatak and Mookherjee (2009) address this question with a somewhat different emphasis, focusing on ex ante
investment incentives by landlord and tenant.
5
For more discussion of this in the case of West Bengal, see Banerjee et al. (2007).
design. The study of compensation and transfers in the face of uneven growth is one of the
most important research areas in development economics.
2.2.4 Conflict. The issue of social conflict in the process of economic development has only
recently begin to receive attention from economists: for example, see the comprehensive
survey by Blattman and Miguel (2009). In principle, one might include all sorts of social
unrest under this rubric, ranging from peaceful demonstrations, processions and strikes to
violent riots and even civil war. In practice, data limitations have restricted study to violent
incidents, such as civil war, guerilla attacks and riots.
The linkages between economic growth and conflict are obviously complex. One connection
is between overall growth (or income levels) and conflict. Collier and Hoeffler (1998, 2004)
and Fearon and Laitin (2003) conducted cross-section studies of civil war across a large
sample of countries. They present correlations showing that per-capita income is negatively
related to conflict. In a study aimed at exploring causal linkages, Miguel, Satyanath and
Sergenti (2004) use annual rainfall shocks as an instrument for economic growth, and support
the findings in sub-Saharan Africa: negative growth shocks appear to raise conflict.
But overall growth is made up of two kinds of changes: one that creates a larger pot to
fight over — and therefore raises conflict — and another that raises the opportunity cost
to fighting, and so lowers it. Whether conflict is positively or negatively related to growth
will therefore depend on the type of growth; specifically, how uneven it is across sectors or
groups. Cross-country studies are too blunt to pick these effects up in any detail, though
again Collier and Hoeffler (1998, 2004) and Fearon and Laitin (2003) led the way with cross-
section studies that connect civil war to the abundance of natural resources. Dube and Vargas
(2009) study civil conflict in Colombia, and obtain more nuanced results: positive shocks
to resources such as oil, the funds from which accrue directly to governments, increase the
payoff from conflict and thereby increases measured violence. On the other hand, positive
shocks to resources such as coffee, which are grown privately, increase the opportunity cost
of engaging conflict: violence declines.
Complicating the task of drawing connections from growth to conflict is the fact that many
conflicts appear to be largely ethnic, geographical, and religious in nature, while outright
economic class struggle is relatively rare. Horowitz (1985), a leading researcher in the area of
conflict, observed that “in much of Asia and Africa, it is only modest hyperbole to assert that
the Marxian prophecy has had an ethnic fulfillment.” Indeed, using measures developed in
my previous work with Joan Esteban (Esteban and Ray (1994)), Montalvo and Reynal-Querol
(2005) have observed that ethnic and religious polarization is significantly and positively
related to conflict. Recent models of ethnic conflict include Caselli and Coleman (2006),
Esteban and Ray (2008, 2009) and Robinson (2001).
There is no contradiction between the use of noneconomic markers in conflict and the
view that conflict may be driven by economic forces. Noneconomic divisions can be and
frequently are used to obtain economic or political gains by violent means, often through
exclusion. Nothing dictates that the groups must be economically distinct. Indeed, one might
argue the contrary: if two groups are very similar economically, it is more likely that they
will intrude on each other’s turf: the motives for exclusion and resource-grabbing — and
therefore for violence — may be even higher.
One classic example of a noneconomic marker is religion, as epitomized by the recurrent
episodes of Hindu-Muslim violence in India. In recent research with Anirban Mitra, we
explore the possibility that growth in the incomes of one religious group might spur
conflict, as the other group finds itself threatened and attempts to nullify that growth.
The econometric analysis we conduct is eminently supportive of this hypothesis. It is also
complemented by several case studies that suggest that one religious group may react to the
economic success of another group by excluding that group through conflict (see Mitra and
Ray, 2010, and the references therein).
This sort of research raises further questions. Which kinds of economic growth reduce
conflict, and which kinds exacerbate it? What is the relationship between inequality and
conflict? How is class conflict related to ethnic conflict, and how can we separate the roles
played by economic incentives and by long-standing distrust and hostility? These empirical
and theoretical questions go well beyond the narrow confines of economics. For our present
purposes, the bottom line is that while long-term, sustained growth should finally tame
conflict, there is every possibility that medium-term uneven growth, among groups or
sectors, could inflame it.
2.3 Uneven Development: Final Remarks
I end this chapter with two broad remarks that pertain to uneven growth and its implications.
The first has to do with institutions. The second has to do with psychological responses to
economic change, particularly to uneven growth.
Of course, institutional success and failure are at the heart of the issues discussed here. After
all, institutions are the structures that safeguard or hinder property rights, compensation
mechanisms, legal structures, political participation, land reform, fiscal structure and a
whole lot besides. But “institutions” is perhaps too broad a term — and too glibly invoked
— to be particularly useful. It may be more productive, instead, to think of the political
and economic agents that have control over various aspects of institutional creation and
implementation.
Sokoloff and Engerman (2000) and following them, Acemoglu, Johnson and Robinson (2002)
and others have emphasized situations of stagnation in which the losers (or potential losers)
control institutions. Losers defend an old system — likely one born under a colonial umbrella
— and so impede progress. But in the developing world of rapid and uneven growth, it
is entirely possible as well that the winners are the ones granted institutional control. This
can be deeply problematic as well, for unless their horizons are long enough to encompass
the possibility that they, too, could be losers in succeeding rounds, they will block all
redistributions that spread the growth process to other sectors. The reaction to such lack
of compensation is, often, violent conflict. It will be of great importance to build a useful
taxonomy of institutional performance (and reactions to such performance) depending on
who has control.
Finally, I remark on psychological reactions to economic change. The fact that uneven growth
can both raise our ambitions, and unleash our frustrations, should be fundamental to our
understanding of economic development. The aspirations of an individual are typically
generated and conditioned by the experiences of others in that individual’s “cognitive
neighborhood”. Such conditioning will affect a plethora of socio-economic outcomes: the
rate of savings, the decision to migrate, fertility choices, technology adoption, the adherence
to norms, the choice of ethnic or religious identity, the work ethic, the strength of mutual
insurance motives, or the collective decision to engage in conflict. Yet the models that
economists write down and apply are blissfully devoid of such social influences on behavior.
The great bulk of individuals living in developing countries are constantly exposed to
economic standards that comprehensively exceed their own. This is true, not just of the poor,
but also of the middle class and the rich. After all, developing countries exist in a global
context in which the wealth of the first world is constantly on display. It is unclear whether
this constant exposure to higher living standards is conducive to a greater commitment to
productive economic betterment, or to despair and frustration (Ray (1998, Sections 3.3.2
and 7.2.4), Ray (2006), Genicot and Ray (2009)).6 A small, “reachable” economic gap may
encourage investments and effort, while a large and persistent gap may stifle it. The raising
of aspirations is a two-edged sword. There is scope for a self-sustaining failure of aspirations
and economic outcomes, just as there is for ever-progressive growth in them.
A research agenda in development economics should not ignore this component of human
behavior. In particular, economists should: 1) build databases from questionnaires that
include socially determinants of behavior, such as aspirations, in ways that allow separating
the informational effect from the hope and desire effect; 2) conduct more research on
inequality tolerance and evolution; 3) study redistributive policy with these ideas in
mind: not necessarily from a normative perspective,7 but from a positive political-economy
perspective.
I write this while on sabbatical in India: a desperately poor country that nevertheless has
an active nuclear program and has sent an unmanned probe to the moon. Just as individual
aspirations drive the dynamics of accumulation within countries, there is a role, too, for
national aspirations, driven by inter-country disparities in consumption and wealth, and
the effect of such aspirations on public policy as well as the international distribution of
income. Even the simplest growth model that exhibits the usual features of convexity in its
technology and budget constraints could give rise in the end to a world distribution that is
bipolar. For instance, countries in the middle of that distribution would tend to accumulate
faster, be more dynamic and take more risks as they see the possibility of full catch-up within
a generation or less. One might expect the greatest degree of “country mobility” in this range.
In contrast, societies that are far away from the economic frontier may see economic growth
— exponential or otherwise — as too limited and too long-term an instrument, leading to
6
These remarks are also related to Duflo’s (2006) hypothesis that “being poor almost certainly affects the way
people think and decide”.
7
Indeed, once we open the door to socially determined preferences, normative economics suffers a setback. That
is not to say that welfare economics is impossible in this context, but several aspects of it need to be looked at
afresh. I do not emphasize this important research area here.
a failure, as it were, of “international aspirations”. Groups within these societies may well
resort to other methods of potential economic gain, such as rent-seeking or conflict.
CHAPTER 3
The Calibration Game
The simple model of convergence also has to be put through enormous contortions to fit the
most essential development facts regarding per-capita income across countries. This is the
point of the current section.
3.1 Some Basic Facts
Low per capita incomes are an important feature of economic underdevelopment—perhaps

the most important feature—and there is little doubt that the distribution of income across
the world’s nations is extraordinarily skewed.
The World Development Report (see, e.g., World Bank [2003]) contains estimates for all
countries, converted to a common currency. By this yardstick, the world produced
approximately $32 trillion of output in 2001. A little less than $6 trillion of this — less
than 20% — came from low- and middle-income developing countries (around 85% of the
world’s population). Switzerland, one of the world’s richest countries, enjoyed a per capita
income close to 400 times that of Ethiopia, one of the world’s poorest.
A serious discrepancy arises from the fact that prices for many goods in all countries are not
appropriately reflected in exchange rates. This is only natural for goods and services that are
not internationally traded. The International Comparison Program publishes PPP estimates
of income, and under these the differences are still huge, but no longer of the order of 500:1.
Over the period 1960–2000, the richest 5% of the world’s nations averaged a per capita
income (PPP) that was about twenty-nine times the corresponding figure for the poorest
5%. As Parente and Prescott [2000] quite correctly observed, interstate disparities within the
United States do not even come close to these international figures. In 2000, the richest state
in the United States was Connecticut and the poorest was Mississippi, and the ratio of per
capita incomes worked out to slightly less than 2!
Of course, the fact that the richest 5% of countries bear approximately the same ratio of
incomes (relative to the poorest 5%) does not suggest that the entire world distribution
of incomes has remained stationary. Of greatest interest — a recent financial crisis
notwithstanding — is the meteoric rise of the East Asian economies: Japan, Korea, Taiwan,
22 The Calibration Game
Singapore, Hong Kong, Thailand, Malaysia, Indonesia, and (commencing somewhat later)
China. Over the period 1965–90, the per capita incomes of the aforementioned eight East
Asian economies (excluding China) increased at an annual rate of 5.5%. Over 1990–1999,
the pace slowed somewhat, especially in Japan, but averaged well over 3% per year for the
remainder.1
Impressive as these rates are, they are dwarfed by China’s phenomenal performance.
Between 1980 and 1990, China’s per capita income grew at an annual rate of 8.6%. The
corresponding figure for the 1990s is even higher: around 9.6%.
In contrast, much of Latin America languished during the 1980s. After relatively high rates
of economic expansion in the two preceding decades, growth slowed to a crawl, and in many
cases there was no growth at all. Morley’s [1995] study observed that in Latin America, per
capita income fell by 11% during the 1980s, and only Chile and Colombia had a significantly
higher per capita income in 1990 than they did in 1980. It is certainly true that such figures
should be treated cautiously, given the extreme problems of accurate GNP measurement in
high-inflation countries, but they illustrate the situation well enough. With some notable
exceptions (such as Chile, 5.7%, and Argentina, 3.6%), annual per-capita growth in incomes
continues to be extremely slow for Latin America in the 1990s, though these rates did turn
positive through most of the region.
Similarly, much of Africa stagnated or declined over the 1980s. Countries such as Nigeria
and Tanzania experienced substantial declines of per capita income, whereas countries such
as Kenya and Uganda barely grew in per capita terms. Notable turnarounds in the 1990s
have occurred in both directions, with alarming declines in countries such as the Congo,
Rwanda and Burundi, and substantial progress in Uganda.
Looking at the overall picture once more without naming countries, one can get a good
sense of the world income distribution by looking at mobility matrices, an idea first applied
to countries by Danny Quah. I’ve constructed one such matrix using 132 countries over the
period 1980–2000; see Figure 3.1.
Each row and column in this matrix is per-capita income relative to world per-capita income.
The rows represent these ratios in 1980; the columns the corresponding ratios in 2000. The
cell entries represent percentages of countries in each row-column combinatioon, the rows
adding up to 100 each. So, for instance, 88% of the countries that earned less than than a
quarter of world per-capita income in 1980 continued to do just that in 2000.
Clearly, while there is no evidence that very poor countries are doomed to eternal poverty,
there is some indication that both very low and very high incomes are extremely sticky.
Middle-income countries have far greater mobility than either the poorest or the richest
countries. For instance, countries in category 1 (between half the world average and the
world average) in 1980 moved away to “right” and “left”: less than half of them remained
where they were in 1980. In stark contrast to this, fully 88% of the poorest countries
(category 1/4) in 1980 remained where they were, and none of them went above the world
average by 2000. Likewise, another 88% of the richest countries in 1980 stayed right where
1
To appreciate how high these rates of growth really are, note that for the entire data set of 102 countries studied
by Parente and Prescott, per capita growth averaged 1.9% per year over the period 1960–85.
The Calibration Game 23
Average to Twice
Half to Average
Quarter or less
Quarter to Half
Twice or more
Quarter or less 76 12 12
Quarter to Half 31 10 7
Half to Average 9 20 46 26
Average to Twice 24 24
Twice or more 5 95
Figure 3.1. The Income Mobility of Countries, 1980–2000.

(Thanks Liz!)
they were. This is interesting because it suggests that although everything is possible (in
principle), a history of underdevelopment or extreme poverty puts countries at a tremendous
disadvantage.
There is actually a bit more to Figure 3.1 than lack of mobility at the extremes. Look at the
next-to-poorest category (those with incomes between one-quarter and one-half of the world
average in 1962). Almost half of them dropped to an even lower category. Thus it is not only
the lowest-income countries that might be caught in a very difficult situation. In general, at
low levels of income, the overall tendency seems to be movement in the downward direction.
My book contains a corresponding mobility matrix for 1963–1984, with very similar findings.
To summarize, then, we have the following observations.
(1) Over the period 1960–2000, the relative distribution of world income appears to have been
quite stable. The richest 5% of the world’s nations averaged a level of per capita income
that was about 29 times the corresponding figure for the poorest 5%. By any standards,
this disparity is staggering, and especially so when we remember that we are talking about
incomes that have been corrected for purchasing power parity.
(2) The fact that the overall distribution has remained stationary does not mean that there
has been little movement of countries within the world distribution. Of particular interest
in the 1980s is the rise of the East Asian economies and the languishing of other economies,
particularly those of sub-Saharan Africa and Latin America. Diverse growth experiences
such as these can change the economic composition of the world in the space of a few
decades. Nonetheless, a single explanation for this diversity remains elusive.
(3) The observation that several countries have changed relative positions suggests that there
are no ultimate traps to development. At the same time, a history of wealth or poverty does
seem to partly foretell future developments. The mobility of countries appears to be highest
somewhere in the middle of the wealth distribution, whereas a history of underdevelopment
or extreme poverty appears to put countries at a disadvantage.
(4) That history matters in this way is an observation that requires a careful explanation.
Poor countries do seem to have some advantages. They can use, relatively free of charge,
technologies that are developed by their richer counterparts. Scarce capital in these countries
should display a higher rate of profit, because of the law of diminishing returns. They can
learn from mistakes that their predecessors have made. In this way differences across
countries should iron themselves out over the longer run. Thus the observation that history
matters in maintaining persistent differences needs more of a justification than might be
obvious at first glance.
One can see different attempts to reconcile the failure of convergence with the traditional
theory:
(A) Hardline View. Don’t abandon the traditional aggregative theory but seek reasons for
productivity and other controls to be systematically different across countries. Conditional
on those controls, attempt to establish convergence.
(B) Multiplicity View. Abandon the convergence argument. Argue that the same fundamen-
tals can progress in very different directions depending on initial conditions.
(C) Interactive View. Argue that the world is one interactive system and cannot be split up
into several growth models (with or without convergence) running side by side.
In these notes we shall spend some time with each of these views.
3.2 The Parente-Prescott and Lucas Calibrations
Normalize labor to 1, and consider the following aggregative production function:

yt = At ktθ ,
where At captures some exogenous growth of TFP. It will be convenient to bestow on At the
exponential form
At = A(1 + γ)(1−θ)t ,
so that (1 + γ)(1−θ)t − 1 can be thought of as the growth rate of TFP. We have written things
in the slightly ugly form so as to simplify the expressions later.
One way to think about it is to interpret γ as the growth rate of labor productivity. That
would translate into output productivity by a power of 1 − θ.
Write the capital accumulation equation:
kt+1 = (1 − δ)kt + xt ,
where δ is the depreciation rate and x is the flow of fresh investment. By the savings–
investment equality and the Solow assumption that savings is proportional to per-capita
income, we see that
xt = syt ,
where s is the savings rate. We can combine all these equations and proceed as follows:
Define kt∗ ≡ kt /(1 + γ)t ; then the capital-accumulation equation implies that
xt
(1 + γ)t+1 kt+1
∗
= (1 − δ)(1 + γ)t kt∗ + ∗ kt∗ ,
kt
and dividing through by (1 + γ)t , we may conclude that
yt ∗
(1 + γ)kt+1
∗
= (1 − δ) + s k.
kt t
Now yt /kt simply equals A/kt∗ 1−θ , so that kt∗ simply converges to k∗ , where k∗ is given by
A
(1 − δ) + s = 1 + γ,
k∗ 1−θ
or !1/(1−θ)
sA
k =
∗
.
γ+δ
It follows that kt converges to the path
!1/(1−θ)
sA
(1 + γ)t
,
γ+δ
so that yt converges to the path
!θ/(1−θ)
sA
A(1 + γ) t
.
γ+δ
We may therefore conclude that in steady state, the effect of varying the savings rate only
depends on the ratio θ/(1 − θ).
But this has small effects. We know that in a Cobb-Douglas world of perfect competition,
θ is a good proxy for the share of capital in national income. Lucas (1990) uses an estimate
for that share of around 0.4, so that the ratio in question is around 2/3. This means that
a doubling of the savings rate — a huge increase — will only raise steady state per-capita
incomes by a factor of 22/3 , which is around a 60% increase. This comes nowhere close to
the inequalities we see around us.
Parente and Prescott (2000) impute around 70% to labor income and 5% to land, which leaves
them with a capital share of 25%. With that our required ratio is even lower: θ/(1 − θ) = 1/3.
That means that a doubling of the savings rate only translates into a 25% variation in per-
capita income.
And moreover, the savings rates in the richest countries are nowhere close to double that
of their poor counterparts. In 1993, the industrialized countries averaged a savings rate of
19.4%. The LDCs actually had a higher savings rate during that period — 23.3% — while
even Africa had a savings rate of 18.8%.
Notice that TFP differentials in productivity give us a better chance to explain differentials:
whereas across two countries 1 and 2,
y1 s1 θ/(1−θ)
= ,
y2 s2
for technical levels the difference is more amplified:
y1 A1 1/(1−θ)
= .
y2 A2
When θ is 1/3, the savings difference translate into income differences as the square root,
while for technology differences the ratio is taken to the power 1.5. So, for instance, a
doubling of the technology difference “explains” a difference of close to three times in per-
capita output. This is a bit closer to what we see, and it is small wonder, then, that those who
are wedded to Solow-type convergence models have been inclined to focus on technological
differences.
Actually, the above calibrations, which are due to Parente and Prescott (2000) (see also
Mankiw, Romer and Weil (1992)) can be given a sharper and more immediate expresson by
simply using the production function and no more. This is the route taken in Lucas (1990).
Once again begin with the Cobb-Douglas production function
y = Akθ
(note, no time subscripts, I won’t even need to do any growth theory). Use the competitive
condition to assert that the rate of return to capital is given by
r = Aθkθ−1 ,
or equivalently
r = θA1/θ y(θ−1)/θ .
Once again take the share of capital to be 1/3; then θ = 1/3, so that across two countries 1
and 2,
!2
r1 y2
= .
r2 y1
This yields absurd numbers. If the per-capita income in the US is 15 times larger than that of
India, the rate of return on capital in India should be over 200 times higher! Even if the share
of capital in production is taken to be 0.4 (used by Lucas), the ratio in the rates of return
should be close to 60, also plainly absurd.
There are alternative routes out of this dilemma.
3.2.1 Differences in Human Capital. Of course, we should get the immediate alternative
out of the way first, which is that labor qualities in the two countries are not the same. So
differences in per-capita income are not the same as differences in income per “effective
capita”. For instance, Anne Krueger (1968) attempts to compare US and Indian workers
by looking at information on each country’s mix of workers (by age, education and sector)
and combining this with (US-based) estimates of how these factors affect productivity (as
measured by relative earnings).
Krueger obtains an overall ratio of one US worker = approx. 5 Indian workers. This means
that the ratio income per effective capita is 3, but this too generates a rate of return differential
between 5 (if capital’s share is 40%) and 9 (if that share is set lower at 1/3). This difference is
also “too large”, and there is still a lot left to explain.
[Update on Krueger and per-capita income differentials: Heston, Summers and Aten (2002)
argue that in 1990 the PPP income differences were perhaps 11:1. Banerjee and Duflo (2004)
adjust the Krueger estimates of relative worker productivity to about 10:3. This leaves us
effectively in the same place: the adjusted ratio is then about 3.2, which creates the same
differentials in the rates of return as in the previous paragraph.]
3.2.2 Differences in TFP. Now let’s turn our attention to technological differences. Let
us look at the implicit TFP ratios needed if we were to equalize rates of return in the two
countries and maintain the requirement that per-(effective) capita income ratios are around
3. Use subscripts I for India and U for the US; then the equality of the two rates of return to
capital demands that
AI yθ−1
I = AU yθ−1
U ,
so that
yU AU 1/(1−θ) AU 1.5
= ' ,
yI AI AI
provided that the share of capital is around a third. So this means that
AU
' 32/3 = 2.08.
AI
It is hard to get a feel for whether this is a large difference or a small one. One way of looking
at it: if the US and India put in the same amounts of capital and quality-corrected labor into
production, the US will produce twice as much as India. This may be a tall order.
Nother way: Lucas’s view is that this difference is attributable to an externality created by
human capital. Suppose that the externality is proportional to ha , where a is some coefficient
and h is the human capital endowment per capita. Then
!a
AU hU
= .
AI hI
Lucas estimates a at around 0.36, using Denision’s productivity comparisons within the
United States over 1909 and 1958, and combining them with human capital endowments
over the same period. Because 50.36 ' 1.8, this takes care of the problem as far as Lucas is
concerned.
3.2.3 Misallocation of Capital. Another way to think about it is to generate the produc-
tivity differences from the misallocation of capital in a disaggregated model. Banerjee and
Duflo (2004) adopt this approach, but there is an interesting tension here. To generate
serious misallocation problems, one must presume that the marginal product of capital is
substantially different across small and large firms. But this means that capital has high
curvature in production, so that one must choose correspondingly smaller values of θ.
Assuming that capital is misallocated cannot provide a ready fix on this problem.
That said, credit constraints and consequent misallocation of resources may well be
important.
3.2.4 The Share of Capital. One way out is to somehow enlarge the share of capital, and
in this way the value of θ. Parente and Prescott (2000, p. 44–55) discuss this route in some
detail, by considering intangible forms of capital and the possibility that physical capital is
grossly mismeasured, but these adjustments are just not enough.
3.2.5 Government Failure. One view is that governments might expropriate new in-
vestors, while existing investors (who may be unproductive) are overprotected. This is
a view in which incumbent elites are not necessarily the best business hands, yet they are
in a position to control the entrance of others more efficient than they are. This is related
to political-economy arguments made by Engerman and Sokoloff and Acemoglu-Johnson-
Robinson that we will discuss later in the course.
Parente and Prescott consider a variant of this point of view, in which they regard the
government as intervening excessively and thus lowering productivity.
Another sort of government failure may arise from the lack of intervention, such as
intervention to protect property rights. Certain types of long-run investment may then
not be made (see Besley, Bandiera, or Goldstein-Udry). Or there may be various free-rider
problems in joint production, as also overexploitation of the commons.
3.3 Summing Up
Convergence relies on diminishing returns to “capital”. If this is our assumed starting point,
the share of capital in national income does give us rough estimates of the concavity of
production in capital. The problem is that the resulting concavity understates observed
variation in cross-country income by orders of magnitude. Huge variations in the savings
rate do not change world income by much. For instance, doubling the savings rate leads
to a change in steady state income by a factor of 1.25, which is inadequate to explain an
observed range of around 20:1 (PPP). Indeed, as Lucas (1990) observes, the discrepancy
actually appears in a more primitive way, at the level of the production function (even
without the attendant steady state theory). For the same simple production function to fit
the data on per-capita income differences, a poor country would have to have enormously
higher rates of return to capital; say, 60 times higher if it is one-fifteenth as rich. This is
implausible. And so begins the hunt for other factors that might explain the difference.
What did we not control for, but should have?
This is the kind of mindset that you will take on board if you get on the convergence boat. The
Solow benchmark of convergence must be tested against the empirical evidence on world
income distributions, savings rates, or rates of return to capital. The two will usually fail to
agree. Then we look for the missing variables that will bridge Solow (or some close variant
thereof) to the data. Thus it is not uncommon to find economists “explaining” inter-country
variation by stating that one country is more corrupt than another, or more democratic, or
is imbued with some particularly hardworking cultural ethic.
With careful economists such as the ones I have cited here, the argument is conducted far
more responsibly. “Human capital” is often used as a first port of call: might differences
here account for observed cross-country variation? The rest is usually attributed to that
familiar black box: “technological differences”. As one might imagine, that slot can be
filled in a variety of ways: externalities arising from human capital, incomplete diffusion of
technology, excessive government intervention, within-country misallocation of resources,
take your pick. All of these — and more — are interesting candidates, but by now we have
wandered far from the original convergence model, and if at all that model still continues
to illuminate, it is by way of occasional return to the recalibration exercise, after choosing
plausible specifications for each of these potential explanations.
The Solow model and its immediate variants don’t do a bad job. In the right hands,
the model serves as a quick and ready fix on the world, and it organizes a search for
possible explanations. Taken with the right grain of salt, and viewed as a first pass, such
an exercise can be immensely useful. At another level, playing this game too seriously
reveals a particular world-view. It suggests a fundamental belief that the world economy
is ultimately a great leveller, and that if the levelling is not taking place we must search
for that explanation in parameters that are somehow structurally rooted in a society. These
parameters cause economic growth, or the lack of it.
To be sure, the factors identified in these calibration exercises do go hand in hand with
underdevelopment. So do bad nutrition, high mortality rates, or lack of access to sanitation,
safe water and housing. Yet there is no ultimate causal chain: many of these features go hand
in hand with low income in self-reinforcing interplay. By the same token, corruption, culture,
procreation and politics are all up for serious cross-examination: just because “cultural
factors” (for instance) seems more weighty an “explanation” does not permit us to assign it
the status of a truly exogenous variable.
In other words, the convergence predicted by technologically diminishing returns to inputs

should not blind us to the possibility of nonconvergent behavior when all variables are
treated as they should be — as variables that potentially make for underdevelopment, but
also as variables that are profoundly affected by the development process.
This leads to a different way of asking the development question, one that is not grounded
in any presumption of convergence. Quite unlike the convergence hypothesis, the starting
presumption is distinct: two economies with the same fundamentals can move apart along
very different paths. Several factors might lead to such divergence, among them various
processes of cumulative causation, or poverty-traps, or initial histories that determined —
at least to some degree — the future that followed.
CHAPTER 4
Expectations and Multiple Equilibrium
4.1 Complementarities
Let n be the number of players and A1 , . . . , An be n action sets, one for each player. Suppose
that the sets are ordered by “≥”. For each player i there is a payoff function πi : A → R,
where A is the product of the actions sets. Say that this game exhibits complementarities if
whenever a−i ≥ a0−i , then
arg max π(ai , a−i ) ≥ arg max π(ai , a0−i ).
ai ai
It will suffice for the purpose of these notes to provide a simplified and more special
description. Suppose that a set of individuals all have access to some set of actions A,
taken to be a subset of the real line. Denote by a a generic action, ai the action taken by
individual i, and by mi the average of all actions other than the one taken by i.
Assume that the payoff function is given by πi (a, m) for each individual i, where a denotes
his action and m denotes the average action taken by everybody else. Then it is easy to see
that there are complementarities in our more general sense if for all i,
(4.1) πi (a, m) − πi (a0 , m) is increasing in m
whenever a > a0 are two actions in the set A.
Notice the difference between complementarities and positive externalities. The former
change the marginal gain to taking an action while the latter affects payoff levels. Changes in
the marginal gain are compatible with payoff levels going in either direction.
As we shall see, Pareto-ordered outcomes are typical of these situations (though they won’t
necessarily happen).
4.2 Some Examples
4.2.1 Qwerty. There are two technologies; call them [Q]werty and [D]vorak. There are
many individuals, each of whom employs a single Q-trained secretary or a single D-trained
secretary. The cost of installing each technology is the same, but the cost expended on a
32 Expectations and Multiple Equilibrium
secretary is a decreasing function of the number of other people using the same secretary
type. [More secretarial schools exist for that type.] This is a situation of complementarities.
The same goes for technologies such as PCs and Macs, in which the benefits from adopting
the technology depend positively on the number of other users (networking).
4.2.2 Infrastructure. A railroad is used for transporting products from the interior to the
ports. People are indexed on [0, 1], and person i gets a benefit B(i) from being able to use the
railroad. The cost of railroad use is declining in the number of users: c(n), where n is the
number of users and c0 (n) < 0. This is a situation of complementarities.
4.2.3 Finance. A thicker financial market caused by lots of people putting their money
in financial assets can create the possibilities of greater diversification. So at the margin, it
becomes easier for an individual investor to invest.
4.2.4 Capital Deepening. Greater roundaboutness in production increases the produc-

tivity of capital, the scale of aggregate production, and in this way the final demand
for individual machine varieties. This may in turn justify the greater roundaboutness of
production.
4.2.5 Social Capital. High rural-urban migration can destroy social capital back in rural
areas. In turn, that destruction can increase the pace of rural-urban migration.
4.2.6 Discrimination. Individuals discriminated against may not invest in human capital,
perpetuating that discrimination.
4.2.7 Currency Crises. Apart from the fundamentals of holding or selling a currency,
there is a strong incentive to sell if other individuals are selling. This forms the basis of a
class of currency-crisis theories based on complementarities.
4.2.8 Endogenous Growth. Economy-wide investment raises the return to individual

investment, thus potentially generating a sustainable growth path.
4.2.9 Social Norms. Sometimes, social norms can change a Prisoner’s Dilemma to a
coordination game. Examples: spitting in public, throwing garbage on the streets, or
engaging in tariff wars. Sometimes repeated interactions can imitate the same outcome
(though with many agents this is hard).
4.3 Complementarities and Development
The ordinary view of capitalist development is that it inflicts negative externalities: pollution,
greed and so on. This is certainly true. But there is an important sense in which the capitalist
investment process creates severe complementarities (whether the underlying externalities
are negative or positive; they could be either).
Expectations and Multiple Equilibrium 33
For instance, a firm that prides itself on quality and fair dealing will induce its competitors
to take the same actions simply to maintain business competitiveness, and could spark off a
quality race (the same applies to research and innovation, or indeed, low prices). Note that
the underlying externalities are negative but that we have a case of complementarities in the
appropriate action space.
In another context, the combined actions of several firms can (a) lower infrastructural costs,
(b) create demand for each others’ products, both directly and (c) by creating higher incomes;
and can (d) enable the creation of new products or the startup of some other productive
activity by making inputs available. These are complementarities, too, in the sense that
these actions of “investment” increase the incentive for other firms to ‘invest” as well. This
time the externalities are positive.
So complementarities can exist both with positive and with negative externalities. Indeed,
the medieval guild system was designed to avoid some of the complementaries which had
negative externalities attached to them — e.g., a new improvement had to be sanctioned
by existing guild members. As you can imagine, such an outcome cannot be stable unless
severe punishments were available for offenders, and such punishments themselves became
weaker as the guild system died away, provoking others to leave the guilds. [Thus the guild
story is also one of complementarities!]
It is very important to understand that while the distinction between positive and neg-
ative externalities is important in understanding the normative properties of a particular
equilibrium, the distinction between complementarities and what might be called anti-
complementarities (reverse the movement of (4.1) in others’ actions) is essential in under-
standing the possible variety in the development experience. Complementarities create the
possibility of multiple equilibria, so that we might argue that countries — or societies — are
in different equilibria though there is nothing intrinsically different between them.
4.4 Multiple Equilibrium
4.4.1 Complementarities through Demand. How do complementarities manifest them-

selves in multiple equilibria? One way to begin exploring this is through the standard
general equilibrium model. Understand, first, that some amount of limited competition is
necessary. If the price at which an output is being sold (or an input being bought) is fixed,
then the agent can inflict no externality on another — he internalizes these externalities
through the price. But if the price drops as you sell more of a product (or the input price
raises), you are creating a gain for another agent that you fail to internalize. Thus perfect
competition may be at odds with the multiple equilibrium story that creates a set of Pareto-
dominating equilibria (of course, multiple equilibria that are Pareto-undominated are clearly
possible by classical considerations).
The route envisaged by Rosenstein-Rodan was two-fold. Investments everywhere create a
climate for more investment: (a) directly, via intersectoral changes in price, and (b) indirectly
via the generation of incomes. The parable of the shoe-factory emphasized (b). This is the
Rosenstein-Rodan story as extended by Murphy, Shleifer and Vishny [1989].
4.4.2 Pecuniary Externalities. Before I get into the Murphy-Shleifer-Vishny model it is

important to observe that the Rosenstein-Rodan view is one of pecuniary externalities, in
which one sort of change (investment) provokes another (investment somewhere else) via a
change in prices. Such externalities are to be contrasted with what one might call technological
externalities, in which there is a direct effect that has nothing to do with prices. (Using a PC
when lots of other people use a PC is a networking effect that is a technological externality.
Of course, one could also tell a pecuniary story in which a big market for PCs drives down
their prices and therefore makes it more attractive to buy one.)
Pecuniary externalities are natural. Tibor Scitovsky (1954), in a famous article, argued that
they are orders of magnitude more compelling than technological externalities. Indeed,
because of Rosenstein-Rodan and Hirschman, pecuniary externalities are viewed as the
fundamental process underlying development (or the lack of it).
But percuniary externalities are hard to model using competitive markets. Basically the
first fundamental theorem of welfare economics stands in the way. That theorem rules
out Pareto-ranked equilibria whenever there are no technological externalities. Let us very
quickly review the theorem in the context of a CRS production model.
Suppose that each person i hasPa utility function ui and a vector of endowments x(i). The
total endowment vector is x ≡ i x(i). Denote by c the vector of final consumption goods: it
is allocated among the population of all individuals with person i getting c(i).
There is a production technology T to convert x into c. Thus T is a set of feasible (c, x)
output-input pairs. We don’t make any particular assumption on T.
Note that production may generate positive profits so we have to distribute them among
the agents. Let θ(i) be the share of aggregate profits π that accrue to agent i.
Finally, there is a price vector p for the final goods and P w for the endowments. Say that
(p∗ , w∗ , c ∗ (i)) is a competitive equilibrium if, defining c∗ ≡ i c∗ (i), we have profit maximization:
π ≡ p∗ c∗ − w∗ x ≥ p∗ c0 − w∗ x0 for all (c0 , x0 ) ∈ T,
and utility maximization subject to budget constraints:
c∗ (i) maximizes u(c(i)) on {c(i)|p∗ c(i) ≤ w∗ x(i) + θ(i)π}.
Proposition 4.1. A competitive equilibrium is Pareto optimal, so in particular there can be no
Pareto-ranked equilibria.
Proof. Suppose not. Then there exists an alternative vector of outputs c (feasible, so that
(c, x) ∈ T) and an allocation c(i) of it such that u(c( j)) ≥ u(c∗ (j)) for all j, with strict inequality
for some j. But then by utility maximization, we must have
p∗ c(j) ≥ p∗ c∗ (j) for all j, with strict inequality for some j,
and adding over all j, we must conclude that
p∗ c > p∗ c∗ .
Substracting the common term w∗ x from both sides of this inequality, we contradict the profit
maximization property.
This is why we employ some imperfect competition in the models of pecuniary externalities
that follow.
4.4.3 Model 1. The Profit Externality. There is a continuum of sectors indexed by q ∈ [0, 1].
The utility function for a consumer is
Z 1
ln x(q)dq.
0
With this utility function, when consumer income is given by y, an amount of y is spent on
every good q.
Now, normalize wage rate to unity; then y = π + L, where π is profits and L is the labor
endowment of a typical agent.
Now suppose that each sector has two technologies, a cottage technology which is freely
available without any setup cost, and an industrialized technology, which requires a setup
cost. In the former, assume that one unit of labor produces one unit of output. In the latter,
assume one unit of labor produces α units of output, where α > 1. But there is a setup cost,
which we denote by F > 0.
Now the cottage sector is competitive (while demand for each good is unitary elastic), so it
follows that if there is industrialization in some sector the price will be set at the limit price,
equal to one. Thus the profit from industrialization is given by
y α−1
(4.2) y− −F= y − F ≡ ay − F
α α
The point, therefore, is that a larger y is more conducive to industrialization. To complete
the circle, more industrialization is also conducive to a larger value of y. To see this, notice
that if a fraction n of the sectors do industrialize, then profits per firm are
π(n) = ay − F
so that aggregate income y(n) is given by
y(n) = nπ(n) + L = n[ay(n) − F] + L,
or equivalently
L − nF
(4.3) y(n) = .
1 − an
Notice that
(aL − F)/(1 − an)
y0 (n) =
1 − an
and that
aL − F
(4.4) π(n) = ay(n) − F =
1 − an
so that combining these two equations we get
π(n)
(4.5) y0 (n) =
1 − an
Note that (4.5) exhibits a multiplier-like quality because of the externality. Extra profits
create more than their own weight in income, while higher income in turn can spur
industrialization. This is a classic case of complementarities.
But it isn’t so classic in one respect. Oddly enough, despite the complementarity, there can
only be a single equilibrium in this model. If F < aL, then this means that at y(0) = L it is
worth industrializing, so everyone will. On the other hand, if F > aL, then from equation (4.4)
it is not worth industrializing even when income is all the way at y(1) (set n = 1 in (4.4) and
evaluate π(1)).
[To be sure, when F = aL there are, in fact, a continuum of equilibria but they are all
equivalent in that they generate the same level of national income.]
One possible “explanation” for uniqueness is that the model is too simple: there isn’t enough
heterogeneity among the firms. Let’s satisfy ourselves that this has nothing to do with it.
Suppose that the fixed costs vary smoothly across sectors, all the way from zero to infinity.
Order the sectors so that F(0) = 0, F(1) = ∞, and F(i) is smoothly increasing. Then if n sectors
invest, it must be the interval of firms [0, n] that’s doing the investing, and the following
zero-profit condition must hold:
(4.6) ay(n) − F(n) = 0,

where
Z n
(4.7) y(n) = π(i)di + L,
0
and
(4.8) π(i) = ay(n) − F(i)
for each i. Let A(n) denote the average value of all the fixed costs on [0, n]; then combining
(4.7) and (4.8) we see that
y(n) = any(n) − nA(n) + L
or
L − nA(n)
y(n) = ,
1 − an
and using this information in (4.6), and moving terms around, we see that
(4.9) [1 − an]F(n) + anA(n) = aL
is the fully reduced-form zero-profit condition. Can there be more than one solution in n
to this condition? Differentiating the LHS of (4.9) shows that the derivative is [1 − an]F0 (n),
which is positive. There cannot be more than one solution.
Why does the complementarity have no effect? It does not because the externality is
generated by the payoff of the firm. At all points for which the marginal payoff is positive
(and so is the externality), the firm pushes ahead and produces more. It does not internalize
the externality but it does not need to — the privately profitable and the socially profitable
moves coincide. Likewise for the case in which profits are negative. There is a cutback which
enhances both private and social outcomes at the same time. So even though there is an
externality, the actions fully internalize the externality, as it were.
One lesson from all this is that the Rosenstein-Rodan intuition needs to be examined more
carefully. The source of the complementarity must be something other than private profit
alone.
The main idea in what follows — and this is perhaps the lesson of the Murphy-Shleifer-
Vishny exercise — is that (to obtain multiplicity) one needs an externality source which isn’t
effectively internalized by the externality provider. For instance, if there are taxes on output
(with proceeds given, say, lump-sum to the economy), then the profit-maximum for a firm
will not correspond to the point at which the marginal externality washes out. Likewise for
the example that we discuss below.
4.4.4 Model 2. The Wage Externality. Suppose that industrialization not only benefits
the industrializing firm, it also benefits workers in that firm in the form of higher wages.
These higher wages need not mean higher utility: they could be compensation for a higher
disutility of labor. Or they may be Shapiro-Stiglitz type wages designed to prevent shirking
in a hard-to-monitor activity. Or there may be political and economic pressures to keep
wages above some stipulated minimum in the organized industrial sector.
Let the (additive) wage premium be v, so that the wage to be paid is 1 + v. Now the
monopolists’s profit from industrializing in any particular sector is
1+v
(4.10) π= y− y − F(1 + v)
α
when the total demand for that sector is given by y.
This specification can generate multiple equilibria. If there is no industrialization, then
y = L. This is self-justifying if profits, evaluated at this level of income, are nonpositive.
That is, using (4.10), we have the condition
1+v

(4.11) L 1− − F(1 + v) ≤ 0.
α
On the other hand, to see if one can have an equilibrium in which all industrialize, one must
have the condition that profits are nonnegative when aggregate income (and hence demand)
is evaluated at the full-industrialization point. Aggregate income is given by
1+v

y(1) = Π + L(1 + v) = 1 − y(1) − F(1 + v) + L(1 + v),
α
or
(4.12) y(1) = α(L − F).
[This is an intuitive expression, since the RHS corresponds to the total output produced.] So
profits are nonnegative at the full industrialization point if
1+v

1− α(L − F) − F(1 + v) ≤ 0,
α
and rewriting this, we obtain the condition

1+v

(4.13) L 1− − F ≥ 0.
α
Now (4.11) and (4.13) can easily be compared to see why multiplicity is possible.
A similar exercise can be carried out using the heterogeneous-cost variant. Make the same
assumptions on the function F(i) as before. Then the equilibrium condition is given by
1+v

(4.14) y(n) 1 − = F(n)(1 + v).
α
Notice that Z n Z n
vny(n)
y(n) = π(i)di + L + v F(i)di + ,
0 0 α
where
1+v
π(i) = y(n) − y(n) − (1 + v)F(i),
α
and combining these last two equations we see that
Z n Z n
1+v vny(n)

y(n) = ny(n) 1 − + L − (1 + v) F(i)di + v F(i)di +
α 0 0 α
Z n
(4.15) = any(n) + L − F(i)di
0
Employing (4.15) in (4.14), we see that
Rn
L − 0 F(i)di 1+v

1− = F(n)(1 + v),
1 − an α
or
Z n
(1 + v)(1 − an)
(4.16) L− F(i)di = F(n) ·
0 [1 + (1 + v)/α]
[Notice that (4.16) reduces to (4.9) when v = 0, so that there is no possibility of multiple
equilibrium when v = 0.]
You can differentiate both sides of this condition, just as we did before, to provide restrictions
for which multiple equilibria are possible.
Model 3. Separation of Costs and Benefits
Another example of the general principle discussed earlier, but we don’t do this in class; of minor
importance.
If the setup costs and the later revenues are separated in time, then the latter will drive
demand, while industrializers care about revenues net of setup costs. This will be sufficient
to drive a wedge between private and socially optimal decision-making.
There is a problem though. For the economy to go through this deferred industrialization
process, the interest rate must rise to encourage consumption postponement. For the
industrializers, this brings up the opportunity cost of investing in industrialization.
If no one industrializes, then income is (L, L) and β = 1
1+r solves out for the interest rate. So
the condition for no industrialization is simply
(4.17) βaL − F ≤ 0.
On the other hand, if full industrialzation occurs, current output is L − F and in the next
period it is αL. This must be rationalized by the Euler equation for aggregate consumption.
If u denotes the indirect utility of current consumption of the composite, then solve
max u(c1 ) + βu(c2 ),
subject to
1
c1 + c2 = present value of income,
1+r
so that
1 u0 (c2 ) u0 (αL)
=β 0 =β 0 .
1+r u (c1 ) u (L − F)
It follows that the condition for full industrialization is
u0 (αL)
(4.18) βaαL 0 − F ≥ 0.
u (L − F)
Are (4.17) and (4.18) compatible? Yes, if the interest rate does not rise by too much. [Actually,
with the logarithmic specification employed by Murphy, Shleifer and Vishny [1989], this does
not work, but if the indirect utility function of composite consumption has higher elasticity
than logarithmic, (4.17) and (4.18) can be compatible.]
4.4.5 Complementarity through the Production Process. “The division of labor is

limited by the size of the market.” Thus spake Adam Smith, and he spake well. A large
market size encourages investments that may be complicated and costly, but have immensely
high productivity. The deeper insight of Allyn Young was to note that the converse is also
true: the “division of labor”, or production roundaboutness, also determines the size of the
market. In this sense, roundaboutness begets roundaboutness.
The example that we consider has to do with the provision of intermediate inputs that
are required in the production of final output in the economy. One feature of economic
development is the creation and use of increasingly sophisticated methods of production,
often characterized by their “roundaboutness.” Almost any productive activity can serve
as an example. Let’s take construction. In developing countries, construction is a pretty
labor-intensive activity. The area is cleared by hand, rubble is removed in small baskets
carried by hand, cement is often mixed at the site and carried by hand, and walls are put
up brick by brick. In industrialized economies, each of these tasks has been automated:
cranes are used for clearing and prefabricated walls are erected at the site. Each automated
instrument, in turn, is produced through a complicated activity: think about how cranes
and prefabricated walls are themselves produced. Thus the final production of a house is
reduced to a large series of automated steps, each of a high degree of sophistication and
requiring the provision of many intermediate inputs.
These sophisticated inputs can be extremely costly to produce if they going to be sold in tiny
markets. The manufacture and sale of cranes requires that there be a fairly large demand
for cranes in construction, and so it is with prefabricated walls. Otherwise it is simply not
worth setting up separate plants to manufacture these items. In other words, intermediate
inputs are often produced under conditions of increasing returns to scale.
At the same time, the provision of intermediate inputs, and the consequent roundaboutness
of production, can have very productive consequences, because production not only benefits
from scale, it also benefits from the variety of inputs that are employed. To see this in concrete
terms, suppose that output is produced using a constant returns to scale technology that
includes as inputs intermediate goods as well as labor. If the quantity of labor as well as of all
existing varieties of intermediate inputs is doubled, then output doubles: this is just a feature
of constant returns to scale. This notion suggests that if the production budget is doubled,
output doubles too. Not so, simply because a doubling of all inputs is only one option under
a doubling of the budget. It is also possible to expand the variety of intermediate inputs that
are used in production. The option to expand variety leads to a situation where output more
than doubles: with input variety, increasing returns to scale is built in provided that the
underlying production function exhibits constant returns. It follows that the productivity
of the economy depends on the scale and richness of operations of the intermediate goods
sector.
To formalize this, suppose that an intermediate good — which we loosely call capital —- is
“produced” by means of several intermediate inputs (machines). Machines are indexed by
i: they appear on a continuum [0, ∞). The more machines are used, the more variety there
is in the production process, but all machines are used ina symmetric way. the easiest way
to capture this is to suppose that
"Z n #σ/(σ−1)
1− σ1
(4.19) X= x(i) di
0
where n is an index of the variety of machines or the roundaboutness of production, and
σ > 1 proxies the degree of substitution across machines. This is like a CES production
function where the elasticity of substitution exceeds 1 (so that no one machine is necessary
in the production process).
Suppose that capital is produced by means of a “budget” K which can be used — at a
normalized price of one — to produce intermediate inputs. Then by strict concavity of the
production function in (4.19) and the symmetry of the problem, the budget would be equally
divided among available machines, so that K = xn would solve for x, the quantity of each
machine used in the production process. Consequently,
(4.20) X = {n (K/n)(σ−1)/σ }σ/(σ−1) = n1/(σ−1) K.
It is in this sense that variety can be equated with total factor productivity.
We are now in a position to see how this leads to multiple equilibria. Suppose that the
economy is “poor” and exhibits a low demand for the final product. This situation means
that intermediate production cannot occur at an economically viable scale, which means
that the prices of intermediate goods are high. Consequently, firms substitute away from
intermediate goods to the use of raw labor. This lowers productivity because of the argument
in the previous paragraph and generates low income in the economy. Low income in turn
generates a low demand for the final good, and the vicious cycle is complete. The other side
of the coin is a virtuous circle. High demand for the final consumption good increases the
demand for intermediates, and because these intermediates are produced under conditions
of increasing returns to scale, prices of intermediates fall. Falling prices encourage a further
substitution away from labor to intermediates, which raises the productivity of the economy.
Incomes rise as a consequence and so does demand, completing the virtuous circle.
A formalization follows, based on Ciccone and Matsuyama [1996].
The idea of this exercise is to endogenize the degree of roundaboutness in production (notice
in passing that the same kind of analysis can be carried out using variety in consumption).
To this end, we suppose that each of the machine sectors must be set up (with a fixed cost of S,
denominated in terms of labor) before production can commence. Thereafter production of
machines takes place under constant returns to scale (see description below). Each machine
sector is run by a monopolist, but since different machines can — in principle — be very
close substitutes in production, there are limitations on his pricing behavior. This is what
we begin by exploring.
Imagine that varieties [0, n] are in force. Then producers of the final good will demand a
quantity x(i) of machine i, chosen to
Z n ! Z n
σ/(σ−1)
max F { x(i) (σ−1)/σ
di} , L − wL − p(i)x(i)di.
0 0
Of course, the input i is of measure 0 but the following calculation can easily be justified by
thinking of this as an approximation for a large but finite number of input varieties. The
necessary and sufficient first-order condition is
Z n
σ σ−1
FX { x(i)(σ−1)/σ di}1/(σ−1) x(i)−1/σ = p(i),
σ−1 0 σ
or
FX X1/σ x(i)−1/σ = p(i),
or
FσX X
(4.21) x(i) = .
p(i)σ
FσX X
[Note: p(i)x(i) = p(i)σ−1
, so demand is elastic.]
Now turn attention to machine producers. Suppose that machines are produced using labor
alone, and that a units of labor are needed to produce one unit of a machine. Then the
marginal cost of producing y units of a machine is given by way, where w is the wage rate.
Using (4.21),we can conclude that producers of the intermediate good will choose p(i) to
FσX X FσX X
max − wa ,
p(i)σ−1 p(i)σ
and (because each machine is of measure 0), this is equivalent to the problem
1 wa
max − .
p(i) p(i)σ−1 p(i)σ
You should check that the first-order conditions characterizing this problem are necessary
and sufficient. They are
(1 − σ)p(i)−σ + waσp(i)−σ−1 = 0,
or
waσ
p(i) =.
σ−1
To make things easier to write, choose units of labor so that a = 1 − (1/σ); then
(4.22) p(i) = w for all i.
Before we go on to determine varieties, think of what it will now cost to buy one unit of the
composite capital X, given that this price determination process is in place. Recall that the
composite is
"Z n #σ/(σ−1)
X = x(i) (σ−1)/σ
di
0
σ/(σ−1)
= n x,
because the same amount of every machine is bought given the pricing rule (11.14). So the
effective price of X — call it P — is given by
cost of buying X pnx w
P= = σ/(σ−1) = 1/(σ−1) .
X n x n
Thus P/w — the effective relative factor price — is equal to n1/(1−σ) , which is declining in n.
FX (X,L)X
So if F is CRS and we denote by α ≡ F(X,L) the factor share of X in production, then
(4.23) α = α(n1/(1−σ) ) ≡ A(n).

If the elasticity of substitution between X and L exceeds unity, A(n) is an increasing function
(it is flat if F is Cobb-Douglas).
Now we return to the problem of determining equilibrium variety. Denoting the value of
final output by Y, the operating profit for a producer of intermediates is given by
px αY
π = (p − aw)x = px(1 − a) = = ,
σ σn
where the second equality follows from the pricing rule p = w, the third equality from the
choice of labor units, and the last inequality from the fact that αY = npx by definition of
factor share. Thus
A(n) Y
(4.24) π= .
n σ
Equation (4.24) tells us that an increase in variety has three effects on the operating profit of
a typical producer of intermediates:
[1] A larger n increases variety and this decreases the share to each variety (this is the 1/n
term).
[2] A larger n affects the factor share of intermediates, generally in a positive direction (this
is the A(n) term).
[3] A larger n affects final output (and therefore national income).
Now let us go ahead and finish the endogenization of n using the free-entry condition:
A(n) Y
π= = startup costs = wS,
n σ
or
nσ Y wL + pnx w(L + anx) + nπ
= = = .
A(n) wS wS wS
Now observe that L + anx = (T − nS), where T is the total labor endowment in the economy,
and π is simply equal to wS. Using these in the equation above, we conclude that
nσ T
(4.25) = .
A(n) S
This expression tells us that the potential for multiplicity is intimately linked to the behavior
of n/A(n) — to its possible nonmonotonicity in n, to be more accurate. For instance, in the
Cobb-Douglas case, A(n) is a constant as we have already seen, so that there is a unique
level of product variety in the economy. On the other hand, if A(n) increases sharply with
n (at least over some range), then the complementarity is strong and multiple equilibria are
indeed possible.
4.4.6 Complementarity and Finance. The focus in this section is not so much the idea
of multiple equilibria as the notion of how externalities might permeate the development
process through different channels. In this section we study the financial sector.
It is well known that financial deepening is one of the characteristics of the development
process. The introduction of money into a subsistence economy opens up opportunities
for trade that never existed before. Similarly, the expansion of the credit system opens up
new opportunities for investment, and this is taken one step further when a stock market
comes into being. Now, the investment-enhancing effects of financial deepening are well
worth studying, but we concentrate here on a slightly different set of questions: how is the
availability of finance (from end-savers) tied to the extent of financial deepening?
The answer that we would like to explore is that financial deepening offers opportunities for
diversification, and this encourages a greater flow of savings from low-risk (but low-return)
activities to the higher-risk (but higher-return) sectors. To be sure, the greater flow permits,
in its turn, greater deepening of the financial market. This is the phenomenon explored in
this section (we draw on Acemoglu and Zilibotti [1997]).
Notice that if each sector costs nothing to set up, then large amounts of diversification can
be achieved even with small amounts of finance by simply spreading the available finance
arbitrarily thinly over the existing sectors. Thus the idea of setup costs (or more generally, a
nonconvexity) must enter the picture again.
Begin by looking at individual choices. Suppose that financial securities for different
production sectors are indexed by j ∈ [0, 1]. Sector j pays off a return of R if state j occurs.
States are realized with uniform probability on [0, 1].
Notice that our goal is not to examine the determinants of productivity in each sector (as we
did earlier), nor do we ascribe different rates of return to each production sector. The only
goal here is to examine diversification possibilities.
Suppose that a measure n of sectors is “open” at any particular date. Rearrange sectors so
that we can think of these as sectors in [0, n], where 0 ≤ n ≤ 1. Now an agent must assign
his current assets to a savings portfolio. Denote assets by A. The sectors available are all
these (of measure n), and a riskless sector which pays a return factor of r. That is, if an agent
assigns a fraction α of his savings to the risky portfolios, and owns a “density” f ( j) of asset
j, his return in state j is
c( j) ≡ (1 − α)Ar + αAR f ( j) if j ∈ [0, n],
(4.26) ≡ (1 − α)Ar if j < [0, n].
Assume for now that these returns are all consumed (later we outline a model that allows
for this as well as for other possibilities). Assume that the agent is risk-averse and that his
utility is given by the familiar functional:
Z 1
ln(c(j))d j.
0
Substituting in the value of c(j) from (4.26), we see that the agent’s objective is to choose α
and the density f (j) to maximize
Z n
ln [1 − α]Ar + αAR f (j) d j + (1 − n) ln ([1 − α]Ar) ,

(4.27)
0
Rn
subject to the constraints that α ∈ [0, 1], f ( j) = 0 for j > n, and 0
f ( j)d j = 1.
The first thing to note about this maximization problem is that given any value of α, the
density f (j) must be uniform over [0, n]. This equalizes agent returns over all states in [0, n],
and by second-order stochastic dominance the agent must prefer this to any other allocation
over the risky portfolio. Consequently, our maximization problem reduces to the task of
finding α ∈ [0, 1] to maximize
αAR

(4.28) n ln [1 − α]Ar + + (1 − n) ln ([1 − α]Ar) .
n
Write down the first-order conditions for this to solve the problem (and verify the second-
order conditions yourself):
R − rn 1−n
= .
αR
(1 − α)r + n 1 −α
Rearrange this to get
(R − r)n
(4.29) α=
R − rn
Notice that α is increasing in n (simply differentiate (4.29)). This summarizes our intuition
that the greater the amount of financial deepening (as proxied by n), the the greater the
amount of savings coming into the (risky) financial sector.
Now we move away from the individual level to understand some more about the process
of opening new financial sectors. Specifically, we wish to caputure the possibility that some
sectors require certain minimum investment sizes in order to open, and that these can vary
from sector to sector. the easiest way to do this is to suppose that there is a function S(i)
defined on [0, 1] that captures the cost of opening a new sector i. We order sectors (this
will not be at variance with our previous ordering) so that S(i) is increasing — assume it is
smooth, with S(0) = 0.
The macroeconomic equilibrium at any one point of time can now be described.
Assuming that n sectors are open, investment per sector — as we have seen — is given by
the amount
αA R−r
= A,
n R − rn
using (4.29) It follows that the equilibrium financial depth is given by
R−r
(4.30) A = S(n).
R − rn
To be sure, there may be multiple solutions to this equation for the usual reason that there
are multiple equilibria, but we are not going to pay this too much attention. The point that
we wish to emphasize is this — as long as there is no equilibrium that involves maximal depth
(n = 1), every equilibrium is Pareto-inefficient. There are externalities in the process that will
not go away, even if coordination failures are resolved.
To understand why, notice that the opening of a fresh sector confers positive benefits on
all agents, simply because of the diversification possibilities that are involved. However,
individual agents wish to equalize their holdings across sectors. This tendency towards
equalization means that sectors with large setup requirements cannot be accomodated. At
the margin, some accomodation must be beneficial. This is the source of the inefficiency.
To make this point more precisely, imagine a situation in which a social planner chooses not
only α and the portfolio for the representative agent in the economy, but also the value of n.
In other words, such a planner would actually maximize
Z n
ln [1 − α]Ar + αAR f (j) d j + (1 − n) ln ([1 − α]Ar) ,

(4.31)
0
subject to all the constraints described after (4.27), and a suitable choice of n, under the
additional constraint that αA f (j) ≥ S( j) for all j ∈ [0, n].
The solution to this problem is a bit more complicated because of this last constraint. In
particular, equal division of the portfolio cannot be assured. But what we can say is this: if
αA f (j) > S(j)
over any interval of securities, then over that interval we must have equal division of the
portfolio. For if not, a second-order dominance move can be created without violating
any of the constraints. It follows that along some interval [0, n], we have equal division,
while possibly along an additional interval [n, n∗ ], we have portfolio holdings that exactly
compensate for the fixed cost of those sectors. So our maximization problem may be rewritten
as: choose n, n∗ (not less than n) and α ∈ [0, 1] to maximize
Z n∗
(4.32) n ln ([1 − α]Ar + S(n)R) + ln ([1 − α]Ar + S(i)R) di + (1 − n∗ ) ln ([1 − α]Ar) ,
n
subject to the “budget constraint”
Z n∗
(4.33) S(n)n + S(i)di = αA.
n
Set up the Lagrangean
Z n∗
n ln ([1 − α]Ar + S(n)R) + ln ([1 − α]Ar + S(i)R) di + (1 − n∗ ) ln ([1 − α]Ar)
n
" Z n∗ #
+λ αA − S(n)n − S(i)di ,
n
and write down the Kuhn-Tucker conditions (we’ll only need to do so for the derivatives
with respect to n∗ and n):
(4.34) ln ([1 − α]Ar + S(n∗ )R) − ln ([1 − α]Ar) − λS(n∗ ) ≤ 0, with equality if n∗ > n,
and
nS0 (n)R
(4.35) − λnS0 (n) ≤ 0, with equality if n > 0.
[1 − α]Ar + S(n)R
Now the question is: is it ever possible for an equilibrium to satisfy the Kuhn-Tucker
conditions, which, as we know, are necessary for an optimum? [Never mind whether they
are sufficient.] The answer is no. To see this, notice that in an equilibrium, n > 0 (because
S(0) = 0). Consequently, (4.35) holds with equality, so that
R
λ= .
[1 − α]Ar + S(n)R
Therefore, if n = n∗ , (4.34) implies that
RS(n)
(4.36) ln ([1 − α]Ar + S(n)R) − ln ([1 − α]Ar) ≤ .
[1 − α]Ar + S(n)R
But this contradicts the strict concavity of the logarithmic function. We know that ln(y) −
ln(x) > (y − x)/y. Applying this to x = [1 − α]Ar and y = [1 − α]Ar + S(n)R, we may conclude
that
RS(n)
ln ([1 − α]Ar + S(n)R) − ln ([1 − α]Ar) > ,
[1 − α]Ar + S(n)R
which contradicts (4.36).
It follows that we must always have n∗ > n in the social planning problem; that is, some
sectors must be “cross-subsidized”. But no agent wants to take up the burden of cross-
subsidization! Given that other agents are holding unequal portfolios in order to keep a
sector open, a given agent will always want to allocate his portfolio equally among the
sectors. To be sure, if all agents do so, some sectors will have to close down, perpetuating
the inefficiency.
Observe that the financial deepening problem has implications for the pace of development,
as well as its uncertainty. In countries with small asset levels (proxied by A), aggregate
development will be beset by frequent crises, as full insurance (diversification) is not possible.
Extend the model a little bit to recursively define A in each period in one of two ways:
[1] A is the product of savings decisions made by young agents who only earn a wage; the
wage is the marginal product of labor evaluated at the going capital stock of the economy.
Total returns to the financial sector constitute the amount of capital, which is ploughed
into the deterministic production of a final good, under an Inada-style production function.
Agents live two periods. This is the approach taken by Acemoglu-Zilibotti and this model
can be easily embedded into sucha dynamic model.
[2] A is the product of savings decisions made by long-lived agents, in which the log utility
function postulated here must be viewed as the value function of a dynamic program with
underlying utility functions also logarithmic. Income is then the sum of wages and capital
returns at any date. This is a bit more complicated as a dynamic program needs to be solved.
With either of these extensions, A will fluctuate for long periods of time at low levels (if one
starts low) until a certain asset level is reached that permits full or near-full diversification.
At this point the economy settles down to a steady rate of growth. Thus, while the ergodic
behavior of all economies is the same, the transitions can be very different: for a poor
economy, the steady state can be a long time coming.
Notice in conclusion that this is one view of history-dependence. Another has to do with
the possibility that long run behavior may also be different depending on initial conditions.
This is what we take up in coming sections of the course.
CHAPTER 5
History Versus Expectations
So far, we’ve discussed the existence of a virtuous circle of expectations that might lead
communities to Pareto-superior states among multiple potential equilibria. As we have
already noted, it is possible to think of these as generalized coordination games. Once the
game-theoretic analogy is pressed into service, however, certain troubling issues arise.
Rosenstein-Rodan and Hirschman were concerned with multiple equilibria, no question,
but this was only the starting point of their analysis. They took it as self-evident that history
somehow plays a role in pinning down the “starting equilibrium”, and that the role of policy
would be to construct ways of “moving” society from one equilibrium to another.
It is of interest to note that the theory of pure coordination games makes it very difficult,
if not impossible, to ask this question in a well-defined way. Imagine a sequence of such
situations. Suppose that a bad equilibrium outcome has been in force for the previous 100
plays of the game. In what sense does that make it more likely that in the 101st iteration,
the bad equilibrium will occur again? Formulated in this way, the answer must be: in no
sense at all. For as is well-known, all sorts of complicated switches between the good and
bad equilibria can actually constitute a perfect equilibrium of the repeated game. There is
no role for history in such a formulation.
Thus, by and large, this literature ignores a central question raised by Rosenstein-Rodan and
Hirschman: how does an economy “move from a bad to a good equilibrium”? We place this
entire phrase in quotes because it is imprecise: so called transitions from one “equilibrium”
to another must themselves be viewed as the equilibria of some encompassing intertemporal
process.
It appears (unless some convincing psychological model can be found that weds people to
the historical status quo) that some sort of state variable is needed to make sense of the
Rosenstein-Rodan/Hirschman argument. [This will also serve as a tentative introduction to
the next part of the course in which such state variables, such as initial inequalities, are strong
enough to pin down a unique equilibrium.] But the state variable must be introduced in
some “economical way”, that does not rule out — by fiat or by assumption — the possibility
of moving to a new equilibrium.
50 History Versus Expectations
In what follows, we look at some different approaches to the question of equilibrium-picking

when there is potential coordnation failure.
5.1 State Variables and Common Knowledge
The first avenue that we explore is due to Morris and Shin (1998) (see also Carlsson and
Van Damme (1993)). The basic idea is very simple and very powerful. Suppose that a
coordination game is played between several agents — we shall write a concrete one down
in a moment — and that the realization of a state variable affects the payoffs. Suppose that
for certain values of the state variable the play of one action or the other is a dominant
strategy, but for other values of the state there is a coordination problem. If the realization
of the state variable is common knowledge, there is no bite at all in introducing the state
variable: each realization of the state precipitates a different game, and each game is played
separately.
But suppose that there isn’t common knowledge of the realizations. Suppose that we all
see the realized value with a bit of noise. Then I know that you have seen something, but
you may have seen something a bit different from me. Now there isn’t a string of separate
games for each realization of the state variable. The implications can be surprising, as the
following model of financial crises reveals.
Consider a country which has pegged its exchange rate at some value e. (For concreteness,
think of e as the number of dollars required to buy one unit of the domestic currency.) We
shall assume that this exchange rate is unambiguously overvalued, in the following sense:
suppose that there is some random variable θ (the state) on [0, 1] which determines the “true”
exchange rate f (θ) were the currency to be allowed to float at θ. Then e always exceeds f (θ)
for all θ ∈ [0, 1].
But θ also influences the exchange rate: which is to say that f (θ) varies with θ. Arrange
so that f (θ) is strictly increasing in θ. So the idea is that θ is some “fundamental” which
influences the country’s capacity to export or import, or to attract investment; the higher
being θ, the more favorable the climate.
Now there is a bunch of speculators (of total measure 1), each of whom can sell one unit
of the local currency. If they do, they pay a transactions cost t. If the government holds
the peg, the exchange rate stays where it is, and this is the payoff to selling: −t. If the
government abandons the peg, then the speculators make a profit of e − f (θ), so their net
payoff is e − f (θ) − t.
What about the government’s decisions? It has only one decision to make: whether to
abandon or to retain the peg. We assume that it will abandon the peg if the measure of
speculators exceeds a(θ), where a is increasing in θ (that is, if the basic health of the economy
is better, the government is more reluctant to abandon1).
1See Morris and Shin (1998) for a very simple account of how to derive a(θ) from a somewhat more basic starting
point.
History Versus Expectations 51
We will assume that there is some positive value of θ, call it θ, such that below θ the situation
is so bad that the government will abandon the peg anyway.¯ In other words we¯ are assuming
that a(θ) = 0 for θ ∈ [0, θ]. Then it rises but always stays less than one by assumption.
¯
Consider, now, a second threshold for θ which we’ll call θ̄: this is the point above which no
one wants to sell the currency even though she feels that the government will abandon the
peg for sure. In other words, θ̄ solves the equation
(5.1) e − f (θ̄) − t = 0.
We will assume that such a θ̄, strictly less than one, indeed exists. But we also suppose that
there is a gap between θ and θ̄: that θ < θ̄.
¯ ¯
[If there were no such gap, there wouldn’t be a coordination problem to start with.]
Now we are ready to begin our discussion of this model. First assume that the realization of
θ is perfectly observed by all agents, and that this information is common knowledge. Then
there are obviously three cases to consider.
Case 1. θ ≤ θ. In this case, the government will abandon the peg for sure. The economy is
not viable, all¯ speculators must sell, and a currency crisis occurs.
Case 2. θ ≥ θ̄. In this case no speculator will attack the currency, and the peg will hold for
sure.
Case 3. θ < θ < θ̄. Obviously this is the interesting case, in which multiple equilibria obtain.
There is¯ an equilibrium in which no one attacks, and the government maintains the peg.
There is another equilibrium in which everyone attacks and the government abandons the
peg. This is a prototype of the so-called second-generation financial crises models, in which
expectations — over and above fundamentals — play an important role (see Obstfeld (1994,
1996)).
So much for this standard model, variants of which we have already seen in excruciating
detail. Now we drop the common knowledge of realizations (but we will maintain the
assumption of common knowledge of the information structure that I am going to write
down).
Suppose that θ is distributed uniformly on [0, 1]: its value will be known perfectly at the
time the government decides whether or not to hold the peg or to abandon it. Initially,
however, the realization of θ is noisy in the following sense: each individuals sees a signal
x which is distributed uniformly on [θ − , θ + ], for some tiny > 0 (where θ is the true
realization). Conditional on the realization of θ, this additional noise is iid across agents.
Proposition 5.1. There is a unique value of the signal x such that an agent attacks the currency if
x < x∗ and does not attack if x > x∗ .
This is an extraordinary result in the sense that a tiny amount of noise refines the equilibrium
map considerably. Notice that as → 0, we are practically at the common knowledge limit
(or are we? the question of what sort of convergence is taking place is delicate and important
here), yet there is no “zone” of multiple equilibria! The equilibrium is unique.
What is central to the argument is the “infection” created by the lack of common knowledge
(of realizations). To see this, we work through a proof of Proposition 5.1, with some informal
discussion.
Start by looking at the point θ − . Suppose that someone receives a signal x of this value or
less. What is she to conclude? ¯ She doesn’t know what everyone else has seen, but she does
know that the signal is distributed around the truth with support of size 2. This means that
the true realization cannot exceed θ, so that the government will abandon the peg for sure.
¯ that for all
So she will sell. That is, we’ve shown
x ≤ x0 ≡ θ − ,
¯
it is dominant to sell.
Now pick someone who has a signal just bigger than x0 . What does he conclude? Suppose,
for now, he makes the assumption that only those with signals less than x0 are selling; no one
else is. Now what is the chance — given his own signal x — that someone else has received
a signal not exceeding x0 ? To see this, first note that the true θ must lie in [x − , x + ]. For
each such θ the chances that the signal for someone else is below x0 is (1/2)[x0 − (θ − )],
so the overall chances are just these probabilities integrated over all conceivable values of θ,
which yields (1/2)[x0 − (x − )]. So the “infection” spreads: if x is close to x0 , these chances
are close to 1/2. In this region, moreover, it is well known that the government’s threshold
is very low: close to zero sellers (and certainly well less than half the population) will cause
an abandonment of the peg. Knowing this, such an x must sell. Knowing that all with signals
less than x0 must sell, we have deduced something stronger: that some signals above x0 must create
sales as well.
So let us proceed recursively: Suppose we have satisfied ourselves that for some index n,
everyone sells if the signal is no bigger than xn (we already know this for x0 ). We define xn+1
as the largest value of the signal for which people will want to sell, knowing that all below
xn are selling.
This is a simple matter to work out. Fix x ≥ xn , and imagine any θ ∈ [x − , x + ]. For such
θ, everybody with a signal between θ − and xn (such an interval may be empty, of course)
will attack, by the recursive assumption. Because these are the only attackers (also by the
recursive assumption), the government will yield iff
1
[xn − (θ − )] ≥ α(θ),
2
or
θ + 2a(θ) ≤ xn +
So we can define an implicit function h(x, ) such that the above inequality translates into
θ ≤ h(xn , ).
Put another way, the implicit function h(x, ) solves the equation
(5.2) h(x, ) + 2a(h(x, )) = x +

It follows that if our person with signal x were to attack, her expected payoff would be given
by
Z h(xn ,)
1
(5.3) [e − f (θ)]dθ − t.
2 x−
Now retrace the recursion starting all the way from n = 0: we have x0 = θ − . Then
(remembering that a(θ) = 0 for all θ ≤ θ) it is easy to see that (5.3) reduces to ¯
Z¯ θ
1 ¯
[e − f (θ)]dθ − t.
2 x−
For x ' x0 , this is just
Z θ
1 ¯
[e − f (θ)]dθ − t,
2 θ−2
¯
which is certainly strictly positive. So x1 is well-defined, and x1 > x0 .
Now put x1 in place of x0 , and repeat the process. Notice that h is increasing in x, so if we
replace x0 by x1 in (5.3), then, evaluated at x = x1 , the payoff must turn strictly positive.2
So the new x2 , which is the maximal signal for which people will sell under the belief that
everyone less than x1 sells, will be still higher than x1 . And so on: the recursion creates a
strictly increasing sequence {xn }, which converges from below to x∗ , where x∗ solves
Z h(x∗ ,)
1
(5.4) [e − f (θ)]dθ − t = 0.
2 x∗ −
It is very easy to see that there is a unique solution to x∗ defined in this way. In fact, something
stronger can be established:
Claim. If x∗ is some solution to (5.4), and x0 > x∗ , then

Z h(x0 ,)
1
(5.5) [e − f (θ)]dθ − t < 0.
2 x0 −
To prove this, consider any x0 > x∗ . Then two things happen: first, it is easy to see that
h(x0 , ) − x0 < h(x∗ , ) − x∗ ,
so that the support over which integration takes place in (5.4) is narrowed. Moreover, the
stuff inside the integral is also smaller when we move from x∗ to x0 , because f (θ) is increasing.
So the LHS of (5.4) unambiguously falls when we move from x∗ to x0 , and we are done with
the Claim.
To learn a bit more about x∗ , use (5.2) to see that h(x, ) − x + = 2[1 − a(h(x, ))], so that
Z h(x∗ ,) Z h(x∗ ,)
1 1
0= [e − f (θ)]dθ − t = [1 − a(h(x, ))]e − f (θ)dθ − t,
2 x∗ − 2 x∗ −
or Z h(x∗ ,)
1
e− f (θ)dθ − t = a(h(x, ))e
2 x∗ −
2This is on the assumption that the sequence {x } stays bounded below θ̄. This will certainly be the case, see
n
below, so it’s not really an assumption at all.
A comnparison of this equation with (5.1) categorically shows that x∗ is bounded below θ̄
for small .
So there is a unique solution to x∗ and it is below θ̄, which justifies the previous recursive
analysis (see in particular, footnote 2). Notice also that our analysis shows that every
equilibrium must involve attack for signals less than x∗ .
To complete the proof, we must show that no signal above x∗ can ever attack. Suppose,
on the contrary, that in some equilibrium some signal above x∗ finds it profitable to attack.
Take the supremum of all signals under which it is weakly profitable to attack: call this x0 .
Then at x0 it is weakly profitable to attack. Suppose we now entertain a change in belief
by supposing that everybody below x0 attacks for sure; then this cannot change the weak
profitability of attack at x0 . But the profit is
Z h(x0 ,)
1
[e − f (θ)]dθ − t,
2 x0 −
which is nonnegative as we’ve just argued. But this contradicts the Claim.
So we have proved that there is a unique equilibrium to the “perturbed” game, in which a
speculative attack is carried out by an individual if and only if x ≤ x∗ . As → 0, this has
an effect of refining the equilibrium correspondence dramatically. To describe this, calculate
the threshold x∗ as → 0. The easiest way to do this is the “sandwich” inequality:
Z h(x∗ ,)
1
[e − f (h(x , ))][1 − a(h(x , )] ≤
∗ ∗
[e − f (θ)]dθ ≤ [e − f (h(x∗ , ))][1 − a(h(x∗ , )],
2 x∗ −
which is obtained by noting that f (x∗ − ) ≤ f (θ) ≤ f (h(x∗ , )) for all θ ∈ [x∗ − , h(x∗ , )]. Both
sides of the sandwich go to the same limit, because x∗ and h(x∗ , ) — as well as the realization
of the state — all go to a common limit, call it θ∗ . This limit solves the condition
(5.6) [e − f (θ∗ )][1 − a(θ∗ )] = t.
It is obvious that there is a unique solution to (5.6).
Note: At this point be careful when reading Morris-Shin. There is an error in Theorem 2.
See Heinemann (AER 2000) for a correction of this error which agrees with the calculations
provided here.
5.2 Applying Global Games to a Canonical Multiple Equilibrium Model
A canonical multiple equilibrium model can be described as follows. An economy consists

of two regions or sectors, A and B. A total capital (or labor) endowment of K̄ is split at date
0 between the two regions. Denote by K the capital in region B, so that K̄ − K is the capital
stock in region A. Capital invested in Region A yields a fixed rate of return, normalized to
zero. Region B’s rate of return r is taken to depend positively on its capital endowment:
(5.7) r = f (K)
where f is continuous, strictly increasing, and f (0) < 0 < f (K̄).
Per Capita Sectoral Return

Number of People A A' O B B' Number of People
Old New
Figure 5.1. A Depiction of Sectors A and B.
Assume that each individual has just one unit of capital, which she can place in either Sector
A or Sector B. So there are K̄ individuals.
This model may be conveniently summarized by using a simple diagram.
Figure 5.1 shows us the two sectors. In A the rate of return is flat at the (normalized) level
zero. In B, the return to any one individual to participating in a sector depends positively
on the number of individuals already active in that sector (starting below zero and ending
above). To complete the description, we locate the initial allocation of individuals across the
two sectors (given by history). This is given by OA people in A and OB people in B. You
could think of the line segment AB as the total number of people in the economy: as the
allocation of people changes, the only thing that alters is the position of AB but not its length.
Now, the initial allocation of people has been chosen to deliberately illustrate a point. Even
though B is, in principle, “better” than A (if everyone were in B the rate of return would
exceed that of A), at the starting location, the actual rate of return in A exceeds the actual rate
of return in B.
Note that our examples discussed so far fit in quite neatly into this framework. For instance,
the return in B could be interpreted as the wages paid to people who have decided to acquire
a certain set of skills, as a function of the number of individuals who have already acquired such
skills. In this interpretation, if you were to think of the alternative occupation as conferring
no externalities, you might imagine that the rate-of-return line in A as nearly flat, while the
one in B slopes upwards. Or you could think of B as typewriters with the Dvorak system,
or for that matter as alternative chip designs, while A contains the QWERTY typewriters,
or Intel chips. The rate of return is then to be interpreted as the total amount of satisfaction
accruing to a consumer or user, net of the cost of purchase or usage. In all these examples,
it might be useful to think of OA as a large segment, while OB is very small, perhaps zero.
Now we are ready to make this model run. Imagine that, as indicated in Figure 5.1, the rate
of return in A exceeds the corresponding rate in B. Then, as time passes, individuals will
gravitate from B to A. This describes the failure of of an exciting new sector when there is
enough critical mass to keep the momentum going. Matters will end with everybody in A
and nobody in B.
Figure 5.1 also reveals that if there had been sufficient critical mass, matters could have been
entirely different. For instance, suppose that initial history initially put us at the allocation
OA0 and OB0 , where the rates of returns in the two sectors are exactly equal. Then the
slightest additional tilt towards B can spark off an accelerating tempo of beneficial change,
as people switch over to the new technology (or the new product, or a new way of life).
Now here is a different story that works entirely on the basis of expectations. Begin again in
the situation where OA individuals are in A, while the rest, OB, are in B. Now imagine that
for some reason, everybody believes that everybody else will be in B tomorrow. Never mind
where this belief came from. Simply note that if this belief is genuinely held by someone, he
must also believe that B is the sector to be in, because the return there is higher. Consequently,
he will gravitate to B tomorrow. But if everybody thinks the same way, everybody will be
in B, and the seemingly absurd belief is completely justified (provided that the one-time
cost of moving is not too high, of course). Thus, it would seem from this argument that
history plays no role at all. Irrespective of initial conditions, there are only two self-justifying
outcomes that are possible, everyone in A or everyone in B, and that both these outcomes are
always possible, depending only on expectations.3
To examine the global games approach to this model, introduce a state z that affects the rate
of return to capital in region B. Rewrite (5.7) as
(5.8) r = f (K, z),
where f is strictly increasing in both K and z. As in the previous section z is a random

variable which is imperfectly observed with uniform and independent noise on [z − , z + ].
We impose end-point conditions just as in Morris and Shin. Specifically, suppose that there
exist z̄ and z in the support of z such that f (K̄, z) = 0 and f (0, z̄) = 0. The first condition
¯ some bad states region B is unprofitable
means that for ¯ even if all capital is moved to that
sector, and for other states region B is profitable even if there is no capital in that region.
Because the model is canonical and stripped of unnecessary content, it is very easy to set it
up within the Carlsson-Van Damme-Morris-Shin framework. But the model is more subtle.
Recall that in the currency crisis model, the number of attackers only has a binary impact on
returns: either the currency collapses or it does not. Here K (which is akin to the number of
attackers) has a smoother impact on the return r.
Let’s conjecture what the equilibrium threshold (for vanishing noise ) must be before we
move to a more formal analysis. At any signal x, I believe that I am uniformly distributed
(in terms of signal strength) among the population, when the noise is tiny. In particular, I
3There is also a third outcome in which the rates of return in the two sector are exactly equalized, but it is
possible to rule this out on grounds of “stability”.
believe that the probability of having K people or more above me is precisely

K̄ − K
.
K̄
That is, the probability density of this random variable is uniform. It follows that my
expected payoff conditional on being in sector B is
Z K̄
1
f (K, x)dK,
K̄ 0
which must equal 0 at the limit threshold x∗ . That is,
Z K̄
1
(5.9) f (K, x∗ )dK = 0
K̄ 0
must solve for the limit threshold.
Now for the formalities. Proceed as before: suppose that everyone with signal x ≥ x̄ puts
capital into B. What is the best response? To solve for this, note that at every realization of
the true state z, the probability that anyone sees a signal that exceeds x̄ is given by
(z + − x̄) for z ∈ [x̄ − , x̄ + ]

 1

 2

ψ(x̄, z) =  for z < x̄ −

0
for z > x̄ + .

 1

It is obvious that the mass of capital that will move to B under the presumed strategy is just
ψ(x̄, z)K̄. So my expected payoff from moving to B, conditional on my signal x, is given by
Z x+
1
(5.10) f (ψ(x̄, z)K̄, z)dz.
2 x−
Setting this expression equal to 0 yields my best response φ(x̄). Check out the edges. Note
that if I see x = z − , I won’t move regardless of what the others are doing, so that for all x̄
small enough, φ( ¯ x̄) > x̄. Likewise, if I see x = z̄ + , I will move regardless of what the others
are doing, so that for all x̄ large enough, φ(x̄) < x̄.
Next, suppose that x̄ goes up. Then for every z in [x̄ − , x̄ + ], ψ(x̄, z) comes down. To restore
balance in the setting of (5.10) to 0, x must go up. That means that φ(x̄) is increasing. At the
same time, the slope of φ has to be smaller than 1. For if not, the expression in (5.10) must
turn strictly positive.
Now we can obtain a unique equilibrium threshold x∗ (just as in the last section) that solves
Z x∗ + !
1 K̄
(5.11) f [z + − x ], z dz = 0.
∗
2 x∗ − 2
Of course, this x∗ will vary with the extent of the noise. We must pass to the limit as → 0.
To do so, perform the change of variables
K̄
K≡ [z + − x∗ ].
2
Then dK = (K̄/2)dz, and at the end points we have K = 0 and K = K̄. Using all of this in
(5.11), we must conclude that
Z K̄
1 2K

f K, − + x dK = 0.
∗
K̄ 0 K̄
Taking to 0, we get
Z K̄
1
f (K, x∗ )dK = 0
K̄ 0
as correctly conjectured in (5.9).
5.3 Lagged Externalities
The use of a state variable which may change the fortunes of one sector versus another
(irrespective of sectoral membership) at any one point of time is a useful way of thinking about
(relatively) short-run episodes such as financial crises. It may not be as useful when thinking
about a country or region which is stuck in some low-investment trap for long periods of
time. It is much harder to conceive of a state variable that will signal positive profitability for
the high-investment outcome (regardless of the actions of other investors) in any one period.
In what follows, we take two approaches; one that dispenses with this sort of state variables
altogether (Adserà and Ray [1998]; this section) and one that resurrects state variables, but
has them follow a Markov process over time (Frankel and Pauzner [2001]; next section).
Recall the canonical model, with two regions, A and B. A total capital (or labor) endowment
of K̄ is split at date 0 between the two regions. As before, denote by K the capital in region
B, so that K̄ − K is the capital stock in region A. Capital invested in Region A yields a fixed
rate of return, normalized to zero. Region B’s rate of return r = f (K), where f is continuous,
strictly increasing, and f (0) < 0 < f (K̄). There is a continuum of agents, and each agent
owns a single unit of capital.
Capital is free to move between regions but in what follows, we introduce both a switching
cost as well as an intertemporal structure to the externality. Assume that the cost of moving
from B to A is given by a function ĉA (K) and that the cost of moving from A to B is given by
some function ĉB (K). Say that either of these cost functions exhibits congestion if it increases,
at least over some interval, in the capital in the relevant sector. That is, ĉA exhibits congestion
if it is decreasing over some range, while ĉB does so if it is increasing over some range.
Before we describe agent behavior, let us track the relevant prices. Let γ ≡ {r(t), cA (t), cB (t)}∞ t=0
be some point expectation about the path (measurable in time) of returns and relocation costs
in each region. Future returns are discounted in the standard way, using a discount rate ρ.
Denote by V(γ, i, t) the optimal value to an agent in region i, i = A, B, beginning at time t, when
the commonly anticipated path of returns is γ. Then by standard dynamic programming
arguments, an agent in region i will switch sectors at time t if V(γ, i, t) < V(γ, j, t) − c j (t), will
stay if the opposite inequality holds, and will be indifferent if equality holds.
A path γ is an equilibrium if it is generated by the optimal decisions of (almost) all agents in
response to γ.
To discuss the generation of γ, consider now some exogenously given measurable path
{K(t)}∞
t=0
(recall the interpretation that K(t) is the amount of capital in sector B).4 We assume
that there is some lag (however small) in the speed at which external effects induced by
incoming/outgoing factors affect the going rates of return. From this point of view, we
regard the return function f (K) as representing a “long run level” of the rate of return, once
the economy has settled at a certain level of capital K. We assume that at date 0, r(0) is
precisely f (K(0)) (see (5.7)). Thereafter, we introduce an increasing function g, with g(0) = 0,
such that
(5.12) ṙ(t) = g( f (K(t)) − r(t)).

Thus, the rate of return at any date “chases” the “appropriate” rate of return corresponding
to the division of the capital endowment at that date. The specific functional form of g(.)
determines the speed at which returns adjust. In any case, capital owners will get paid r(t)
at date t.
Thus a path of capital allocations {K(t)} generates a path of returns {r(t)} using (5.12), and a
path {cA (t), cB (t)} using the relationships cB (t) = ĉB (K(t)) and cA (t) = ĉA (K̄ − K(t)) for all t.
Several economic situations conform quite naturally to this specification. In models of
search or matching, the productivity of some fixed amount of capital may depend on the
ability of that capital to find partners (with more capital), say, because of minimum scale
requirements in production. This ability, in turn, will depend on the total amount of capital
in the economy (see, e.g., Diamond [1982]). Note that a discontinuous jump in the capital
stock will lead to a smooth intertemporal increase in productivity as long as the process
of “matching partners” takes place in continuous time. Likewise, if one replaces “capital”
by “population” and “rate of return” by “utility”, the concentration of population in a
particular geographical region may provoke large amounts of productive activity and a
variety of goods and services, attracting still more people because of the greater utility to be
had (Krugman [1991b]). Again, the degree of productive activity might react smoothly to
a sudden influx of population (perhaps because the information regarding a larger market
needs time to permeate to all the producers).
Say that an intertemporal equilibrium is exclusively history dependent if the long-run outcome
either equals the initial allocation, or entails migration only to the sector that is initially
profitable. Note that myopic tatonnement has the same properties, though obviously the
exact path may be different. The main similarity is that no room is left for farsighted
expectations.
Proposition 5.2. Assume f (K(0)) , 0. Unless the cost of relocation exhibits congestion, every
equilibrium must be exclusively history dependent, irrespective of the discount rate.
This observation is independent of the magnitude of discounting, and of the degree of

responsiveness of returns (as long as it is not instantaneous). Thus by a minor and reasonable
weakening of one of the assumptions in the literature, we obtain a class of models where
4Note that we do not a priori restrict {K(t)} to be a continuous path, so that self-fulfilling “jumps” are, in principle,
permitted.
expectations are dwarfed by history, where initial conditions determine the final equilibrium.
Of course, if rates of return adjust instantaneously, then expectations-driven equilibria are
possible.
We reiterate: our claim is not that ahistorical equilibria are impossible. But that, in this
class of models, in addition to the intersectoral agglomeration externalities, the “migration
technology”, is crucial to understand the sources of such equilibria. The only way in which
such outcomes can occur is by introducing a cost to postponement; i.e., by making future
relocation costs increase in the stock of settlers in our case.
Proof of Proposition 5.2. Consider the case in which f (K(0)) < 0. The case f (K(0)) > 0 can
be settled by a parallel argument.
Fix any equilibrium γ. We claim that K(t) ≤ K(s) for all t ≥ s, which settles exclusive
history-dependence.
Suppose this is false. Then, indeed, there is some t and s with t > s, and exhibiting the
following features:
a. K(t) > K(s).
b. r(τ) < 0 for all τ ∈ [s, t].
c. Some agent moves to sector B at date s.
d. The moving agent does not return to A until after date t.
Part (a) is simply the negation of our proposition. Part (b) follows from the fact that returns
adjust continuously, and that r(0) < 0. Part (c) follows from the fact that K cannot begin to
climb unless there is movement from A to B. And Part (d) follows from Part (b): there is no
point in someone going to B and coming back while the returns there are strictly negative
throughout.
Now consider our moving agent at date s and have her move at date t instead, following
thereafter her original optimal strategy (by (d), this is possible). Then her deviation return,
discounted by ρ to date s, is given by
(5.13) e−ρ(t−s) [V(γ, B, t) − cB (K(t))],
while along the presumed optimum it is (once again discounted to s)
Z t
V(γ, A, s) = r(τ)dτ + e−ρt−s V(γ, B, t) − cB (K(s))
s
< e −ρ(t−s)
V(γ, B, t) − cB (K(s))
−ρ(t−s)
≤ e [V(γ, B, t) − cB (K(s))]
(5.14) ≤ e−ρ(t−s) [V(γ, B, t) − cB (K(t))],
where the first inequality follows from Part (b), the third from discounting, and the last from
the no-congestion assumption.
But together, (5.13) and (5.14) yield a contradiction.
5.4 State Variables and Time
We now resurrect state variables but in a way that concentrates on the possibly long-term
nature of the coordination problem (as in Adserà-Ray and in contrast to Morris-Shin). (The
material here is based on Frankel and Pauzner [2001].) As in the previous sections, we
assume that in every period there is a coordination game to be played. However, we will
take the return in Sector B to depend, not just on the amount of K in that sector, but also on
some exogenous state variable zt . Thus we write r = f (K, z) for the return in that sector, while
the return in sector A is normalized to zero just as before. Without loss of any generality
assume that f is increasing in z (the realization of the random variable zt at date t).
The idea is that zt is some random variable which is changing over time and is exogenous
to the model at hand. For example, it could be the price of oil, which makes the industrial
sector B relatively less or more attractive than the agricultural sector A.
We assume that an individual receives a chance to move according to some Poisson hazard
rate λ, upon which he can costlessly choose to be in either sector. Let µ denote his discount
rate. We assume that our individual seeks to maximize the present discounted value of
returns. Let {r(t)} be some anticipated time path of returns. Look at the decision of where
an agent chooses to be at any date at which he has the chance to move; say date 0 without
loss of generality. His choice of sector A will yield 0, and his choice of sector B will begin to
yield the path {r(t)}. The point is that we should only look at this path until he has a chance
to move again, because once he has a chance to revise his choices he can costlessly move to
either sector. The density of new moving chances is just λe−λt , so that the agent’s effective
payoff — to be referred to as simply “payoff” henceforth — is given by
Z ∞ h i
λ e−λt e−µt r(t) dt.
0
In short, he will move to B if Z ∞
e−ρt r(t)dt > 0,
0
and to A if this inequality is reversed, where ρ ≡ λ + µ. Note well, before we leave this point,
that the above integral does not represent his “true” lifetime payoff but only that portion of
it that accrues until he has a chance to move again. But this value will form the basis on
which all decisions are made, and so we use the term “payoff” for it.
Our assumption on z is that it follows Brownian motion with no trend. So the increments to
z all look the same, no matter what the starting level. The second assumption is a joint one
on f and z: there are values of z (in its overall support, which you can think of as the entire
real line) so that it pays for everyone to be in Sector B today no matter what happens in the
future, and no matter how many people are in Sector B today. Likewise, there are values of z
so low that the opposite is true: it is worth being in Sector A today no matter what happens
in the future.
The former assumption can be formally expressed as
"Z ∞ #
(5.15) E e f (0, zt )|z0 = z̄ ≥ 0
−ρt
t=0
for some z̄ in the support of the random variable. The latter assumption can be expressed as
"Z ∞ #
(5.16) E e f (K̄, zt )|z0 = z ≤ 0
−ρt
t=0 ¯
for some z in the support of the random variable.
¯
Notice that this model emphatically does not rule out the multiple equilibrium problem. The
assumptions embodied in (5.15) and (5.16) only rule out the problem for extreme and possibly
very low-probability values that zt might take on at some later date. For “most” intermediate
values of z, there should be a genuine coordination problem as before: my behavior will
depend, in principle, on what I anticipate others to do, now and in the future. Yet this model
admits a unique equilibrium:
Proposition 5.3. In the infinite-horizon moving game, there is a uniquely defined function k̂(z) such
that everyone who has a chance to move at any date, chooses sector B when k > k̂(z), chooses Sector
A when k < k̂(z), and is indifferent when k = k̂(z). The specific nature of the function k̂ will depend
on the data of the problem.
The proposition states that despite the potential scope for multiple equilibria, history once
again fully pins down the outcome (but in a completely different way from the previous
exercise). The extreme values that z can conceivably assume somehow serve to pin down
behavior in the intermediate (more likely) zone.
Understanding this proposition teaches us something subtle about the role of beliefs in
equilibrium models, so it is worthwhile to go through the details. To emphasize the
nontriviality of this proposition, it may be worth considering — just for a moment —
the case in which z does not change at all over time (see Figure 5.2). This sort of diagram will
be repeated more than once so let us get used to it. On the horizontal axis are various values
of z; in particular, the threshold values of z̄ and z are clearly marked. On the vertical axis
are different values of the capital stock in Sector ¯B. The maximum such value is obviously
K̄; this is marked as well.
You are to interpret both these values as current values. [Of course, this injunction is irrelevant
for z at the moment because its value is fixed, but it will become relevant later.]
The diagram also contains two downward-sloping lines. The first of these lines — call it
line I — depicts the combinations of (K, z) values such that if Sector B starts with K, and then
gains capital whenever people get a chance to move, the present discounted value of sector B
is zero. Formally, this first line is the collection of all (K, z) pairs such that
"Z ∞ #
E e f (Kt , z)|K0 = K = 0
−ρt 1
t=0
where {Kt1 } is the process in which people move to sector B whenever they have a chance to
move, starting from K.
Notice that as z goes up, this makes returns better for Sector B, so that the threshold value
of K needed to sstisfy the condition goes down. This is why the line is downward-sloping.
_
K
I II
_
_
z z z
Figure 5.2. Boundaries for Multiple Equilibria when z is Fixed.
The second line — call it line II — describes the combinations of (K, z) values such that if
Sector B starts with K, and then loses capital whenever people get a chance to move, the
present discounted value of sector B is zero. Formally, this first line is the collection of all
(K, z) pairs such that
"Z ∞ #
E e f (Kt , z)|K0 = K, z0 = 0
−ρt 2
t=0
where {Kt2 }
is the process in which people move to sector A whenever they have a chance to
move, starting from K.
Line II is downward-sloping for the same reason that line I is. In addition, it should be
obvious that line II must lie to the right of line I, as it imposes zero profits for Sector B on the
basis of more pessimistic expectations about that sector.
Now notice that lines I and II mark out demarcations for unique and multiple equilibria. If
the system finds itself to the right of line II, everybody must move to B when they get the
chance. Likewise, if the system finds itself to the left of line I, everyone must move to A
when they get the chance. But between lines I and II the situation is up for grabs. Because
such a configuration is to the left of line II, it is possible to speculate that if everyone flees
for Sector A, then a particular agent must too (when she gets the chance to move). This is
self-fulfilling. Similarly, because the configuration is to the right of line I, one can sustain the
opposite movement in this zone. Multiple equilibria are not ubiquitous (for extreme values
of z there is only one equilibrium), but they haven’t vanished, not by a long shot.
So much for the case in which the value of z is exogenously given. Now let’s suppose that z
follows a Markov process as described above. Figure 5.3 starts this analysis by drawing the
analogue of line II: the one with pessimistic expectations from the point of view of Sector
_
K
k*(z) k(z,1) k(z,0)
Figure 5.3. Recursive Construction of the Locus k∗ (z).
B. I say “analogue” because the earlier description took z to be given; now we have to
take expectations over changes in z as well. Formally, we look at all pairs of points (k, z) —
forming the locus k(z, 0) — such that
"Z ∞ #
(5.17) E e f (Kt , zt )|K0 = K, z0 = z = 0
−ρt
t=0
where Kt is the “pessimistic process” in which individuals move to A whenever they get the
chance, and zt is the exogenous stochastic process starting from z0 = z.
The locus k(z, 0) is the analogue of line II, and it is downward-sloping for exactly the same
reason. But now matters get interesting just to the left of this locus. The pessimistic
expectation: can it be validated? It can if z is fixed, but now it is possible that tomorrow,
z might increase, taking the whole system into the area right of k(z, 0). In that case we
know that whoever gets a chance to move will not move to A. So expectations that are fully
pessimistic are inconsistent with what we already know about the model, at least with some
probability. Let us modify, then, our process K to say that it pessimistically moves down
(as before) except under the above contingency. But now this makes Sector B a bit more
attractive than it was when constructing the locus k(z, 0). It follows that the new zero-payoff
locus associated with these somewhat brighter expectations — call it k(z, 1) — must be to the
left of k(z, 0) (see Figure 5.3).
This starts an iteration. Even k(z, 1) must be too pessimistic, because the prospects of moving
to B are brighter than those described by k(z, 0). In general, define k(z, n + 1) to be the locus
of all (k, z) points which make people indifferent between Sectors A and B, assuming that
whenever the system is to the right of k(z, n) people will move to B, and will move to A otherwise.
[Because z has a density, we don’t have to worry about what happens if we land right on
_
K
k'(z,0) k'(z,1) k*(z)
Figure 5.4. Recursive Construction of the Locus k∗ (z).
k(z, n).] All these loci are downward-sloping, and they converge to some limit which we
shall call k∗ (z). This limit locus has the following properties:
[1] If the system is to the right of this locus, everyone who has a chance to move at that point
must move to Sector B.
[2] If the system is to the left of this locus, matters are ambiguous. Certainly, if everyone
is expected to go to A whenever the system is to the left of k∗ (z), then they will do so. But
there is no ruling out multiple equilibria yet. It may be that points to the left of k∗ (z) are also
consistent with a move to B.
So we need to continue the analysis, and to do so we start an iteration from the left this time.
The iteration is not going to be entirely symmetric, so pay attention to what follows.
First, we translate the locus k(z, 0) way over to the left. Figure 5.4 illustrates. By our
assumption on extreme values, I can find, in fact, a translate that is so far over that for every
(k, z) on that (translated) locus,
"Z ∞ #
(5.18) E e f (Kt , zt )|K0 = K, z0 = z ≤ 0
−ρt
t=0
where Kt is now the “optimistic process” in which individuals move to B whenever they get
the chance, and zt is the exogenous stochastic process starting from z0 = z. That is, these
values of z are so bad for sector B that despite the harbouring of optimistic expectations
about future moves to Sector B, it is best to stay in Sector A. In this way, we have identified
a sufficient condition for moves to Sector A to always constitute an equilibrium: that (k, z) be
to the left of this translated locus; call it k0 (z, 0).
But now, once again, we know that these optimistic predictions are too optimistic. People
won’t always move to B; in fact, as we have just seen, if the state moves into the extreme
left zone that we have just identified, they will move to A (or stay there, as the case may
be). Thus we can define a fresh collection of (k, z) with the property that starting from each
of these points and entertaining the optimistic expectation except when the state wanders to
the left of k0 (z, 0), state A is weakly better. Figure 5.4 shows the boundary of this set as a
dotted line to the right of k0 (z, 0).
Now iterate as follows: translate the locus k0 (z, 0) — to the right this time — as far as possible,
subject to the condition that no part of it wander out of the dotted line (see Figure 5.4 to
follow this fully). This translation we will call k0 (z, 1). More generally, continue the iteration
as follows: given k0 (z, n), consider the collection of all (k, z) points which make people weakly
prefer Sector A to B, assuming that whenever the system is to the right of k0 (z, n) people will move to
B, and will move to A otherwise. Having done so, define k0 (z, n + 1) to be the furthest translate
of k0 (z, n) subject to the constraint that it fully lies within the set of points defined in the
previous sentence. Let k∗ (z) denote the limit of this recursion.
Appreciate again that this iteration is different from the first one. It has a similar flavor, but
we only restrict ourselves to translates. Nevertheless, we can certainly say that for all points
to the left of k∗ (z), it is a dominant strategy to move to sector A (after all, we have been — if
anything — “conservative” towards A when performing this iteration).
Now we need to look at points on the locus a bit more carefully. Notice that the first limit —
k∗ (z) — certainly has the property that on k∗ (z), one is indifferent between A and B , provided
we assume everyone moves to B on the right of this locus, and towards A on the left of this
locus. But because we have thrown away some points in doing the second iteration, we
cannot say the same of k∗ (z). All we can say is that if we assume everyone moves to B on the
right of k∗ (z), and towards A on the left of k∗ (z), then on the locus people weakly prefer A to
B.
But we can go one step further: there must be some point on k∗ (z) at which individuals must be
indifferent between A and B, provided we assume everyone moves to B on the right of this
locus, and towards A on the left of this locus. The reason is the nature of our iteration: at
each stage we chose the maximal translation to the right that is possible. If strict preference
for A held at all points on k∗ (z), the iteration could not have ended at that stage.5
Call this point of indifference α. Now go back to the locus k∗ (z) and pick its “twin point” β
(see Figure 5.5). That there is a twin point, at eactly the same vertical height, follows from
the fact that the loci k∗ (z) and k∗ (z) are translates of each other.
Our final claim is that α and β must coincide. To see this, recap the following two points:
[i] α is a (k, z) combination at which all movers are indifferent between staying at A and B,
assuming that everybody (who can) will choose A to the left of the locus k∗ (z) and everybody
will choose B to the right of that locus.
5This is a crucial technical point in the proof and establishing it formally needs a bit of care, though the intuition
is very clear.
_
K
k*(z) k*(z)
Figure 5.5. Showing that the Loci k∗ (z) and k∗ (z) Must Coincide.
[ii] β is a (k, z) combination at which all movers are indifferent between staying at A and B,
assuming that everybody (who can) will choose A to the left of the locus k∗ (z) and everybody
will choose B to the right of that locus.
Now recall that z has no trend. But then statements [i] and [ii] must be incompatible if we
insist that α and β are distinct! To see this, simply couple together the events at α and β,
and thereafter. For every path starting from α, there is a twin path starting from β, forever
separated by the same distance in z. And the two share exactly the same probability measure!
But the expectation of returns over the first set of paths is zero. Therefore the expectation
of returns over the second set of paths must be positive, because f is increasing in z. This
proves that α and β cannot be distinct.
But then the entire functions k∗ (z) and k∗ (z) must coincide! Call this common function k̂(z).
We have proved the proposition. It must be that to the left of k̂(z), everyone moves to A,
while to its right everyone moves to B. Multiple equilibria have been eliminated.
CHAPTER 6
History-Dependence: An Introduction
6.1 An Overview
To summarize what we have studied so far: we’ve looked at multiple equilibrium models
in which there are “good” and “bad” equilibria, often Pareto-ranked. The presence of such
multiplicities has often been put forward as an explanation for underdevelopment. Ot at
least, even if it cannot serve as an explanation, it does succeed in pointing out why entirely
different levels of economic outcomes are consistent with the same underlying fundamentals.
As we have seen, the potential difficulty with the multiple-equilibrium explanation is that
we have no theory of how one equilibrium or the other comes about. Indeed, if we run such
models in real time, there is no particular reason why the economic system cannot “jump”
from one equilibrium to another, such jumps helped along by acts of coordination among
the population. At the same time we know that such jumps make no intuitive sense. An
economy locked into many years of one bad equilibrium is not likely to change overnight
(or at any rate, overnight on any one particular night). We have looked at three different
theoretical frameworks that attempt to come to grips with this. The first takes seriously the
common-knowledge presumptions that underlie these models and shows that a departure
from such presumptions can often pin down unique equilibria, typically mediated by some
public signal. The second approach studies lagged external effects. The third approach
returns to public signals in a dynamic context to precipitate a unique equilibrium.
What’s nice about these models is that they will typically generate rapid transitions from
one equilibrium to another. What these models say that we will observe are long periods
spent in a particular equilibrium, and then a series of events that suddenly hurl the economy
into an entirely different configuration. Many social and economic transformations do have
this “logistic” feature.
In what follows, we take a very different — and complementary — path. Now the state
variable will get center stage and we will push multiple equilibrium into the background.
Indeed, we will consider models in which at every date (and for each going value of the
state variable), there will be a unique outcome. But the dynamics of the system will allow for
multiple steady states. Multiple steady states are therefore not to be confused with multiple
70 History-Dependence: An Introduction
equilibrium. There will typically be just one for every initial condition. But different initial
conditions will often map into different steady states.
There are many different state variables that we can study. Each of them is important enough
to have a book written on them. First begin with economics:
1. The capital stock is, of course, the most common example of a “state” in development
economics. It is the central variable of growth theory. Section 6.3 below tells you a little
bit about the Romer model, in which the initial capital stock can have long-term effects on
economic growth (unlike in the convergence models).
2. Infrastructure represents a special case of item [1], but an extremely important special
case that has yet to be studied carefully in the context of history-dependence.
3. Likewise, it would be of great use to have economic theories of culture and corruption,
which then feed back into economic progress (or the lack thereof). Not that models of
economics and culture don’t exist, but once again there is not that much that fits the history-
dependence framework that I emphasize here.
4. Poor societies have legal systems that are limited, if not in terms of laws on the books,
then in terms of actual enforcement of those laws. The same is true of contract enforcement.
Limited enforcement of contracts and laws then feed back on the economic system.
5. Economic inequality: this will be the one topic that we discuss in great detail below. But
stay tuned for future editions in which I plan to take up all the other themes one by one.
Then there is politics:
1. Elites that deny power and therefore cripple subsequent growth.
2. Historical inequalities in (say) the distribution of land that distorts and limits political
participation to demands for redistribution rather than for growth-enhacing investments.
3. the poverty-conflict-poverty cycle, about which we shall have something to say in these
notes.
6.2 Introductory Notes on Economic Inequality
We now turn to a particular — but important — state variable: the disparity in the historical
distribution of assets, or economic inequality. To be sure, inequality is a concept that is of
intrinsic interest, and there is a large literature that has attempted to axiomatize measures
of inequality so that one can keep track of changing income or wealth distributions from a
purely ethical perspective. But there is also the functional aspect of inequality: the interaction
between inequality and other variables of significance in the economy, such as aggregate
output, efficiency, unemployment, and so on. These interactions may take place “within the
period”, so that they are analyzable using a static model, or they be dynamic, in which case
we must use intertemporal models that keep track — among other things — of accumulation
decisions.
History-Dependence: An Introduction 71
Why does inequality matter (not intrinsically, which is a question of ethics, but functionally)?
There are several answers to this question:
[1] Inequality matters because it affects productive potential. The cleanest parable of this has
come to be known as the nutrition-based efficiency wage hypothesis. In poor societies,
wages affect food consumption, and food consumption affects the ability to work. This will
lead to an equilibrium in which the distribution of asset ownership will have economy-wide
effects.
[2] Inequality matters because it affects the ability of individuals to gain access to productive resources.
As long as the distribution of productive ideas or new projects is loosely correlated with the
distribution of wealth, there will be people who have ideas (and abilities), but do not have
access to the wealth needed to put those ideas into real outcomes. This sort of argument
hinges on a missing or imperfect market for capital.
[3] Inequality matters because it affects incentives. There are various aspects of this problem and
we shall address some of them. For instance, what is the connection between an egalitarian
society and incentives? What is the connection between egalitarianism and the extent of
collective action?
[4] Inequality matters for “political-economy” reasons. Under this catch-all are a number of effects
that may be discussed. Inequality may affect voting behavior, leading to different degrees
of redistribution and consequently affecting the incentives to accumulate capital. Inequality
may create social conflict and in this way be related to low output. Finally, inequality (within
an organization) may create efficiency losses, because individuals at different wealth levels
may effectively have different objectives.
[5] Inequality matters because it prevents correct resource allocation on the part of the government.
That is, high inequality may affect the ability of individuals to transmit information
accurately to a policy-maker.
We begin with dynamic models that track the evolution of inequality over generations. The
emphasis will be on items [1] and [2] discussed above.
6.3 A Special Note on Capital Stocks
We now begin our transition to a more serious assessment of initial conditions. Unlike the
more abstract role for history postulated in repeated coordination games, these considera-
tions arise from full-fledged dynamic models in which state variables play a fundamental
role.
I provide a preliminary illustration of this using a variant of a growth model due to Romer
[1986]. In this model, multiplicity of equilibrium paths can exist from the same initial
condition, but the initial conditions play a fundamental role as well.
Suppose that output is produced at each date t according to the equation

(6.1) yt = F kt , k̂t ,
where k̂ is the average economy-wide stock of capital, introduced to express the social
externalities from capital (or the knowledge embodied in capital). The idea is that greater
capital in the society at large has positive spillovers for any one producer. In the sequel we
will take kt = k̂t for all t, but no one individual internalizes this connection in his decision-
making.
Produced output is divided among consumption and investment as follows:
(6.2) yt = ct + at kt ,
where c represents consumption and ak represents the amount of net investment in the
production of new capital. The fact that we express new investment as a fraction a of the
existing capital stock is simply a matter of convenience.
Investment creates growth in the capital stock. We postulate that
kt+1 − kt
(6.3) = G(at ),
kt
so that the rate of growth of the capital stock depends on the “intensity” of investment: on
the investment-capital ratio a.
The objective function takes the familiar form:
∞
X
(6.4) max βt u(ct ),
t=0
where β, the discount factor lies between 0 and 1, and u is a one-period utility function.
The following assumptions will be in force:
(U.1) The utility function u is increasing, continuous and strictly concave on R+ .
(G.1) The function G(a) is continuous, increasing and concave, with G(0) = 0 and sup G(a) ≡
B < ∞.
(F.1) The production function F is continuous, nondecreasing, and increasing and concave in
its first argument. However, when private and social capital move together, we assume that
the (private) marginal product of capital is nondecreasing; that is, F1 (k, k) is nondecreasing
in k.
Also restrict the discount factor β to ensure that the infinite sum of utilities is always well-
defined. I omit the details here.
Fix some initial stock k and a path of “social capital” stocks ẑ ≡ {k̂t }, with k̂0 = k. Notice
that a program (for some individual) can simply be identified with the (private) sequence
z ≡ {kt } (everything else can be recovered from it).Say that such a program z is optimal from k
under ẑ if it maximizes utility (see (6.4)) in the class of all feasible programs. Finally, z is an
equilibrium program if it is optimal from k under z itself.
Using fixed point arguments, it is possible to prove that an equilibrium exists. But of course,
there may be many of them, and this comes from exactly the same considerations that we
have seen earlier in these notes. The fact that an increase in social capital increases private
marginal product means that a complementarity is at work here — higher investments in
the economy as a whole may be self-fulfilling. The point of this section, however, is not to
emphasize the multiplicity but to see how an initial state (proxied in this instance by the
capital stock) — or history — can determine the set of equilibrium paths.
To this end, define an equilibrium program from k to be stationary if kt = k for all t. We may
now state
Proposition 6.1. There exists an equilibrium program from k which is stationary if and only if
(6.5) F1 (k, k)G0 (0) ≤ β−1 − 1.
Proof. Necessity. Here is an outline which can easily be made fully rigorous. The idea is that
if an equilibrium program is stationary, then a tiny sacrifice of consumption today in return
for a larger stationary stream of output in the future is not worth it. Let’s suppose that the
tiny sacrifice of consumption is . Then the current loss is approximately u0 (c). What is the
future gain? The extra capital stock produced is approximately G0 (0), and consequently the
extra stream of output available from tomorrow is approximately F1 (k, k)G0 (0). If this is all
consumed, then the extra stream of utility from tomorrow is approximately u0 (c)F1 (k, k)G0 (0).
Calculate the present value of this extra stream of gains. Our program is an equilibrium
(and therefore an optimum from k under itself), so this deviation should not be worth it.
This yields (6.5).
Sufficiency. Suppose that expectations of social capital are stationary at k̂t = k for all t. We
want to show that the stationary program is optimal under these expectations, starting from
k. let c be the consumption under the stationary program and let {kt } be any other program
(feasible from k under stationary expectations) with associated consumption sequence {ct }.
Then for any date T,

T
X T
X
βt [u(c) − u(ct )] ≥ u0 (c) βt [c − ct ]
t=0 t=0
T
X
= u (c)0
βt [F(k, k) − F(kt , k) + at kt ]
t=0
T
X
≥ u (c)0
βt [F1 (k, k)(k − kt ) + at kt ]
t=0
T
X 1
≥ u0 (c) βt [F1 (k, k)(k − kt ) + (kt+1 − kt )]
G0 (0)
t=0
T
X 1−β 1
≥ u0 (c) βt [ (k − kt ) + 0 (kt+1 − kt )]
βG (0)
0 G (0)
t=0
T
u0 (c) X
= {(βt−1 − βt )(k − kt ) + βt (kt+1 − kt )}
G0 (0)
t=0
T
u0 (c) −1 X
= {(β − βT
)k − (βt−1 kt − βt kt+1 )}
G0 (0)
t=0
u0 (c) T
= {β (kT+1 − k)} ≥ 0,
G0 (0)
which completes the proof.
Proposition 6.1 shows us that stationary equilibria exist under certain conditions, and
moreover, that they cannot exist unless those conditions are met. Notice that the condition
(6.5) combines technological parameters, the discount factor, and the initial capital stock in a
specific way. If you combine this restriction with our assumption that F1 (k, k) ≥ 0, we see
that at certain low levels of the capital stock, it is possible to stagnate forever.
The role of the initial capital stock is made even sharper in the next proposition, which
tells us that if (6.5) is not met — that is, if the initial capital stock is large enough — then
equilibrium programs must exhibit sustained growth.
Proposition 6.2. If
(6.6) F1 (k, k)G0 (0) > β−1 − 1,

then every equilibrium program must exhibit kt → ∞.
Proof. Notice that {kt } must be a nondecreasing sequence. Suppose, contrary to our claim,
that kt → k̄ < ∞ as t → ∞. Consider the sequence of programs {ktn } given by ktn = k(t + n) for
each n and t. Note that the nth sequence has as its starting stock kn , and that for each n, the
program is an equilibrium from kn . By a standard maximum-theorem-style argument that
we do not include here, the pointwise limit of these programs also represents an equilibrium.
But the limit is the stationary sequence {k̄t } given by k̄t = k̄ for all t. By Proposition 6.1 ,
F1 (k̄, k̄)G0 (0) ≤ β−1 − 1.
But k̄ ≥ k, the original initial stock, and given that k fails (9) and F1 (k, k) is nondecreasing, we
have a contradiction.
Thus in the Romer model, if the initial capital stock is large enough, equilibrium programs
must grow. No self-fulfilling expectational argument can create stagnation. However, if
the initial capital stock is low, an expectations trap is possible. In this way history and
expectations interact.
In case you have not noticed this already, notice that Propositions 6.1 and 6.2 are different in
one significant respect. Proposition 6.2 rules out stagnation, while Proposition 6.1 says that
stagnation is one possible outcome, but does not rule out growth. So there is a role here for
expectations, despite the influence of history. With more work, one can show that there is an
even lower (positive) value of the capital stock from which no growth can ever occur, and in
the remaining “intermediate” zone there are multiple equilibria involving both stagnation
(as we’ve already seen) and sustained growth, depending on expectations.
CHAPTER 7
The Dynamics of Inequality
7.1 A Introductory Framework
Every dynamic model of growth and distribution will exhibit some form of the following
equations at the level of the individual unit (person, household, region, country). Think
of time as being divided into discrete periods and adopt the view that each period is the
length of an entire generation. So individuals indexed by different dates belong to different
generations. The first equation is just budget balance:
(7.1) yt = ct + xt ,
where y is the lifetime income or wealth of an individual in generation t, c is her consumption
and x is the bequest she leaves to the next generation.
The next equation describes how bequests translate into lifetime income for the next
generation:
(7.2) yt+1 = f (xt ),
where f is some production function. Such production functions have different interpreta-
tions depending on the context:
1. It may just be a standard production function as used in the theory of growth, say of the
Cobb-Douglas form.
2. If the individual is part of a competitive economy everyone could earn a wage rate w and
then get a return of 1 + r on bequests, so that
f (x) = w + (1 + r)x.
3. The “production function” may simply represent returns to different “occupations”. For
instance, suppose that it costs nothing to keep your child unskilled but costs x̄ to turn her
into a skilled laborer; then
f (x) = w for x < x̄
¯
= w̄ for x > x̄.
78 The Dynamics of Inequality
in this case the production function will be exogenous to an individual dynasty but may
well be endogenous from the point of view of the economy as a whole (because the wages
for skilled and unskilled labor will be determined in equilibrium). These represent the sort
of interactive models that will occupy a lot of our attention.
The last ingredient has to do with preferences. There are four notions that appear in the
literature.
WG. Warm Glow. Generation t has a utility function given by U(ct , xt ). See, e.g., Banerjee
and Newman (1993).
CB. Consumption-Based. Generation t has a utility function given by U(ct , ct+1 ). See, e.g.,
Arrow (1973), Bernheim and Ray (1986).
IB. Income-Based. Generation t has a utility function given by U(ct , yt+1 ). See, e.g., Becker and
Tomes (1979, 1981).
NP. NonPaternalistic. Generation t has a utility function given by U(ct , Vt+1 ), where Vt+1 is
the lifetime utility of generation t + 1. See, e.g., Barro (1978) and Loury (1981).
Notice that WG starts to look really problematic in those cases in which the production
function is endogenous, as in the case of occupational choice. It is hard to cling to
some exogenous utility function that is based on bequests when what those bequests are
accomplishing is endogenous to the system. Thus, while WG and IP are equivalent when f
is exogenous, WG is not really something we will be looking at seriously.
What about CB versus IB? I prefer the latter for two reasons.
First, I should be getting a utility out of my child’s “capability”, measured here by lifetime
income or wealth. Just what she does with it — how much she consumes, how much she gives
away — should not really be my concern. That said, we might sometimes see paternalistic
parents who care more about the child’s consumption rather than the resources that she has
access to. I’m not saying this is a logical rebuttal of CB.
Second, a model based on CB is typically hard. There are all sorts of game-theoretic subtleties
involved here (see Kohlberg (1976) and Bernheim and Ray (1986)). It does not seem sensible
to bring in all those subtleties and still try to address the different questions of inequality
and development that we are more interested in here.
That said, however, why not go the whole hog and choose NP over IB? After all, NB is
completely nonpaternalistic: parents only care about the utility of their children: not the
bequests they leave them, or how much they consume, or indeed how much they earn.
There is some sense in this view and we will try and incorporate it in what follows. But it
has to be realized that the value-function approach can be conceptually problematic. It is well
known that NP is equivalent to the maximization of a utility function defined on an infinite
generational stream of consumptions. Do we really internalize that much? Do parents really
calculate future utilities, or do they use the “capability measure” — income/wealth — as a
convenienet shorthand?
The Dynamics of Inequality 79
There is another reason — but more on this later — why NP may not be the correct
specification.
In any case, even at this level of abstraction, we can glean a few general principles. The most
important of these — at least for the aggregative models that we consider — is the principle
of monotonicity. The easiest way to see this is to suppose that a generation has the utility
function
Eα U(c, Ψ(k, α)),
where c is its consumption, k is its investment, α is some random shock, and Ψ(k, α) is some
mapping. If it is the identity mapping, then Ψ(k, α) = k and this is like warm-glow. But
Ψ could also represent next-period’s consumption, or income, or utility. For now let us
not worry about the endogeneity of these mappings; we will do that when we get to the
equilibrium analysis.
Assume that
[U] The utility function U is differentiable and strictly concave in c, and exhibits comple-
mentarities between c and Ψ: U12 (c, Ψ) ≥ 0.
Renark. Obviously U is not an ordinal property and all of the above can be written as a
suitable supermodularity condition. However, condition [U] is fine enough for what we
want to do.
Proposition 7.1. Assume [U] and suppose that Ψ is nondecreasing in k. Let h be the policy
correspondence that describes the optimal choice of k for each y, subject to the constraint that c = y−k.
Then if y > y0 , k ∈ h(y), and k0 ∈ h(y0 ), it must be the case that k ≥ k0 .
Proof. Suppose this assertion is false for some (y, y0 , k, k0 ) as described in the statement
of the proposition. Then k0 > k. Notice that k0 is feasible for y (because it is feasible under y0 ,
which is smaller), while k is feasible under y0 (because k0 , which is bigger, is feasible under
y0 ). It follows from optimality that
EU(y − k, Ψ(k, α)) ≥ EU(y − k0 , Ψ(k0 , α)),
while
EU(y0 − k0 , Ψ(k0 , α)) ≥ EU(y0 − k, Ψ(k, α)).
Adding these two inequalities and transposing terms, we see that
EU(y − k, Ψ(k, α)) − EU(y0 − k, Ψ(k, α)) ≥ EU(y − k0 , Ψ(k0 , α)) − EU(y0 − k0 , Ψ(k0 , α)).
Now use the assumption that Ψ is nondecreasing in k, plus the complementarities condition,
to conclude that
EU(y − k0 , Ψ(k0 , α)) − EU(y0 − k0 , Ψ(k0 , α)) ≥ EU(y − k0 , Ψ(k, α)) − EU(y0 − k0 , Ψ(k, α)).
Combine these last two inequalities to obtain that
EU(y − k, Ψ(k, α)) − EU(y0 − k, Ψ(k, α)) ≥ EU(y − k0 , Ψ(k, α)) − EU(y0 − k0 , Ψ(k, α)),
and now draw yourself a diagram to see that the strict concavity of U in c is violated by this
last inequality.
This proposition yields the following corollary when there is no uncertainty:

Proposition 7.2. Suppose that there is no uncertainty, that the production function is increasing
in k, and that the optimization problem above is solved repeatedly by successive generations starting
from some initial y. Then the resulting paths of yt and kt must be monotonic in time.
Proof. Suppose not. Then two cases are possible.
Case 1. There exist dates t and s with s ≥ t + 1 such that yt < yt+1 , yt+1 = · · · = ys , and
ys+1 < ys .
Case 2. There exist dates t and s with s ≥ t + 1 such that yt > yt+1 , yt+1 = · · · = ys , and
ys+1 > ys .
It is easy to see that one of these two cases must occur if the proposition is false. We now
obtain a contradiction in Case 1; the same argument holds for Case 2.
In Case 1, we have yt+1 > ys+1 , it follows from f increasing that kt > ks . However yt < ys .
This contradicts Proposition 7.1.
7.2 Inequality and Capital Markets: Noninteraction
There are two approaches to study of evolving inequality. One is to look at ongoing random
shocks. The other is to argue that there are intrinsic market forces which tend to separate
individuals. The first approach typically consists in looking at the ergodic behavior of a
single dynasty, and then equating this to the cross-sectional distribution of the system as a
whole. This is the viewpoint epitomized in several papers: we discuss a variant of Loury
[1981] in these notes.
On the other hand, it should be obvious that the second type of analysis — inequality
arising from market-generated separations across individuals — cannot be conducted by
simply studying the intertemporal behavior of a single agent. Interactions are fundamental.
This is the topic of the next section. Here we study the single-dynasty model.
There is one agent alive at every date, and each agent has a single child, who becomes an
adult in the next “period” (or generation). An agent at date t has access to total resources yt ,
which we may sometimes loosely call income, but is really to be interpreted as the sum of
income and any starting wealth. The agent divides this into consumption (ct ) and bequests
for the next generation (kt ):
yt = ct + kt .
Bequests create starting wealth yt+1 for the next generation. More than interpretation is
possible: these may be financial bequests, or upfront educational investments (the preferred
interpretation by Loury). In any case, yt+1 is not fully pinned down by kt . A random shock
αt+1 ∈ [0, 1] is also assumed to play a role. One might call this the ability of the agent at date
t + 1. Thus we have
yt+1 = f (kt , αt+1 )
where it is obviously reasonable to assume that f is increasing and smooth in the first
argument and that we simply label abilities so that f is also increasing in the second
argument.
We will make further assumptions on f , but at present we only need to make one: that even
for the highest ability, the production function ultimately “flattens out”: limk→∞ f 0 (k, 1) < 1
(where f 0 is the derivative with respect to the first argument). This can be justified at a couple
of levels, but the main reason we do this is convenience: we want everything to be bounded
so that we can analyze it simply. [Why does this assumption guarantee that income and
capital must be bounded? Draw a diagram of the “best-case” production function under
this assumption and satisfy yourself.]
Note that — unlike Loury — we do not assume that f is concave. This allows us to put other
models in perspective (more on this below).
Now turn to preferences. We shall assume that each generation is a dynastic utility
maximizer; that is, it seeks to maximize
u(c) + δEV,
where u is some smooth strictly concave utility function defined on current consumption,
δ ∈ (0, 1) is a disciunt factor, and V is the expected lifetime utility to be experienced by the
child. The expectation is taken before the V because it is assumed that the parent does not
know how the child’s ability will turn around at the time of making the educational bequest.
That is, given some y, a parent tries to maximize, choosing k,
u(y − k) + δEα V f (k, α)

(7.3)
subject to the constraint that 0 ≤ k ≤ y. Notice that the function V is fundamentally
endogenous and that it must solve the well-known functional equation:
V(y) = max u(y − k) + δEα V f (k, α) .

0≤k≤y
That such a consistent V exists can be proved using standard arguments. which we omit
here. It can also be shown that V is continuous in y.1
The beauty of this sort of argument is that it essentially reduces an infinite-dimensional
problem to something very simple: the one-variable maximization problem expressed in
equation (7.3). From this problem we get what is called an (optimal) policy correspondence
Γ(y). This is the set of all k’s which solve (7.3), given the starting value y. Notice that we
have made no assumptions on the curvature of f so in principle this correspondence could
have arbitrary shape. It turns out that the following monotonicity argument must be true:
Proposition 7.3. Suppose that y ≤ y0 and that k ∈ Γ(y) and k0 ∈ Γ(y0 ). Then it must be that k0 ≥ k.
Proof. Verify that all the conditions of Proposition 7.1 are satisfied.
1The argument used to establish the existence, uniqueness and continuity of V is the famous Banach contraction
mapping theorem. See Stokey and Lucas [1989] for how this works with value functions such as those in this
model.
Recall from Proposition 7.2 that this would immediately yield a convergence argument if
there is no uncertainty. Of course, the current model does have uncertainty so this simple
sequence argument cannot be applied. [Try it and see where it fails.] But nevertheless, the
proposition continues to be helpful in this and in other cases, as we shall see in a bit.
Let’s return to the convergence argument in the face of uncertainty. Is there some way
in which we can approach this problem? Here I follow Loury in making the following
additional assumptions on the production function:
Poor Geniuses Exist. f (0, 1) > 0.
Rich Fools Exist. f (k, 0) < k for all k > 0.
I’ve labeled the assumptions so that they are self-explanatory. Just in case: the first
assumption states that even if you don’t invest anything into a high-ability child, she will
end up making some money. The second assumption states that low-ability children drain
resources: whatever you put in, less comes out.
Under these two assumptions, the following important proposition is true:
Proposition 7.4. Let µ0 be an initial probability measure on the set of initial incomes of generation
zero, and denote by µt the distribution of income for generation t induced by the equilibrium behavior
described above. Then there exists a unique measure µ∗ such that µt converges to µ∗ as t → ∞, and
this measure µ∗ is independent of µ0 .
It is important to understand why this result is true, as it leads to a significant insight into
how uncertainty works in these models. To this end, consult Figure 7.1. He we show
diagrammatically how income evolves from generation to generation within a dynasty,
starting with income y0 at date 0. An investment of k0 is made, and this leads to an uncertain
income at the next generation (the little hill on the t = 1 line marks the density of such
incomes). A typical income realization is the level y1 , at which the investment k1 is made,
and the whole process repeats itself. The density conditional on an investment k1 is shown
by the little hill on the t = 1 line, while the thick density with larger support shows all
possible values of y2 two periods hence (that is, conditioning on y0 but not on the specific
realization of y1 ).
Now, in this diagram I have marked out a special interval of incomes I which has the
property that no matter where you start from, there will be a (common) date T such that
these densities (conditional on the starting point) “envelop” the interval I (with probability
uniformly positive. This fact follows from the two assumptions we’ve made above regarding
“poor fools” and “rich geniuses” (though, as we shall see in a bit, the assumptions are not
necessary). The point is that with positive probability and at some uniformly chosen date,
the system will wander into I no matter what the initial conditions were. Intuitively, what
this means is that the system must lose its memory, its history, at some point of time (for
whatever happens with positive probability must happen for sure). That suggests that limit
behavior is independent of history, which indeed is the broad substance of Proposition 7.4.
t=2
I
=0 =1
k1
t=1
y1
=0 =1
k0
t=0
y0
Figure 7.1. A Dynastic Income Path
If you want to know more about the technicalities of this ergodic theorem, look at the
general result (Proposition 7.5) given in Section 7.2.2. Essentialy, the ability of the process to
“communicate” or “mix” no matter what the initial conditions were, is responsible for the
history-independent limit behavior of the system.
Another way to graphically examine the mixing property is to look at the policy functions
generated as a result of dynastic optimization. While these map from current y to the current
choice of k, we shall do this slightly different: the result is shown in Figure 7.2. The diagram
places current income on the horizontal axis, and tomorrow’s income on the vertical axis.
Of course, tomorrow’s income is uncertain, and may therefore be represented by a band
of possible incomes, which boils down to an interval of incomes for each value of today’s
income.
Now, in part (a) of this figure, notice that there are two clear zones. In Zone I, income must
be bounded above by the value YI , which is the biggest intersection (in this region) of the
stochastic correspondence with the 450 line. By monotonicity of the policy correspondence,
it follows that if y ≤ YI , y can never cross the value YI , even under the most optimistic
conjectures regarding the realization of ability shocks. Likewise, there is a lower bound YII
yt+1 yt+1
450 450
yt yt
YI YII YII
Panel a Panel b
Figure 7.2. Mixing
in a different region of the policy correspondence. By the same logic, if y ≥ YII , y can never
fall below the value YII , even under the most pessimistic realization of ability shocks. Notice
that in this diagram, YII exceeds YI , so that there two regions are effectively segregated.
Proposition 7.4 can thus never hold for this scenario.
Now study panel (b) of Figure 7.2. Here, there is an upper trap YII just as we had before,
but there is no “lower trap”: incomes starting below can always wander into the zone to the
right of YII , and must stay there therafter. Thus there is full mixing and the ergodicity result
of Proposition 7.4 is restored.
But are the two assumptions concerning fools and geniuses met? Not really? We have,
reflected in panel (b), a variant of these assumptions, which does just as well. In fact,
convince yourself by drawing an imaginary panel (c), that if YI is well-defined but there is
no YII , that this would be pretty much in line with the assumptions we do have. But now you
also see that the specific form of our assumptions is not really necessary: what we are after
is really the existence of a mixing zone, as described in Figure 7.2’s panel (b), for instance.
7.2.1 Uncertainty, Inequality, and Ergodicity. What, then, is the precise role played by
uncertainty in these models? To understand this, it will be useful to first strip the uncertainty
away altogether, and simply look at a deterministic version.
First of all, suppose that the model is fully convex: that is, in addition to u being strictly
concave, we shall also assume that f is strictly concave (in addition to assuming the endpoint
conditions f 0 (0) = ∞ and f 0 (∞) < 1). Now the optimal policy correspondence is really a
function, and by proposition 7.3 it must be a nondecreasing function. It is therefore easy
enough to see — by iteration — that the sequence of dynastic incomes {y0 , y1 , y2 . . .} must
converge to some steady state y∗ . It is also easy to see (under our assumptions) that y∗ must
be strictly positive.
At this stage we don’t know what y∗ is. We do not even know whether y∗ is unique or not. Of
course, it is unique given y0 , but nothing in what we’ve said so far rules out the possibility
that it might change as y0 changes. But in fact, this cannot happen, and the easiest way to see
this is to use the Euler equation to the solution of the optimization problem, which shows
us that
(7.4) u0 (ct ) = δu0 (ct+1 ) f 0 (kt )
for all t ≥ 0. In the limit, yt → y∗ , so it is trivial to see that kt and ct must converge as well, say
to k∗ and c∗ respectively. Passing to this limit in the Euler equation (7.4), we may conclude
that
(7.5) δ f 0 (k∗ ) = 1,
which, by the way, is the famous formula for the “modified golden rule” in optimal growth
problems. When f is strictly concave, the value of k∗ is uniquely pinned down. There can
be, therefore, no more than one value for limit income, no matter where one starts from.
Convergence occurs to perfect equality (though this may take time).
This brings us to the first role of uncertainty: by creating ongoing shocks (in this case to
ability), it keeps individuals away from perfect equality. With each shock, one might think
of the convergence problem as beginning again, but it is regularly perturbed by ongoing,
further shocks. Thus uncertainty acts as a tool to create inequality in a world of convergence.
Several economists who rely on convergence-based models generally invoke uncertainty to
“explain” inequality. [In fact, Loury’s model assumes that f is strictly concave and is
therefore an example of this.]
Now suppose that we drop the concavity of f . Now (7.4) is still valid as an interior first-order
condition, though we must be careful about checking second-order conditions. The main
point, however, is that (7.5) will now admit several solutions in general. And indeed, there
are now several steady states, depending on initial conditions. With decreasing returns
dispensed with, history is perfectly capable of creating a lock-in effect. If there is inequality
to start with, it may not go away: individuals may remain stuck in different steady states.
This multiplicity in the face of nonconvexities has been known for some time (see, e.g.,
Majumdar and Mitra [1982] and more recently, the first part of Galor and Zeira [1993]).
But the possibility of such historical lock-in (arising from the convex f ) brings us to a second
role for uncertainty, which is its ability to remove such lock-ins. After all, the analysis in the
previous section did not rely at all on the concavity of f . It is quite possible that the “mixing”
condition discussed there holds even when f has the “wrong” curvature. This means that
uncertainty can actually remove lock-ins and restore equality (or at least equality of long-run
opportunity) when in the presence of perfect certainty, such equality would be missing.
This sort of discussion suggests that a study of uncertainty in this context may be misleading
at the same time that it may be illuminating. It may be misleading because even very small
mixing probabiluities lead us back to ergodicity (for instance, assume that everyone has a
small but positive probability of winning the state lottery; then the mixing condition would
be satisfied). But it may be a long time coming. For this reason, an uncertainty-based theory
may hide certain structural features of the model. (For instance, we we’ve seen, there is a
deep difference between the case of concave and convex f , but somehow the presence of
uncertainty pernitted us to provide a mathematically similar treatment.)
There is a final point about the role of uncertainty that might be worth making. It is peculiar
to the noninteractive models that we have been discussing. Notice that with or without
a mixing condition, it is possible to derive limit theorems regarding the distribution of
income, in the following sense: starting from an initial y0 , the resulting distributions of
income converge. The mixing condition simply adds that there will be a common limit to
which the convergence occurs. Without the mixing condition — as in panel (a) of Figure 7.2
— there will be several steady states, but two different limit distributions must have no incomes
in common. [If they did, the mixing condition would be satisfied, and we would not have
two different limits in the first place.]
This is a peculiar and interesting characteristic of the noninteractive model, and also points to
its inadequacy. If one wants to use such models to explain varying different distributions in,
say, different countries with the same underlying fundamentals, one would have to contend
with the uncomfortable prediction that the poorest person in one country must be richer
than the richest person in the other country! The more realistic (but more complicated)
interactive models that study later will set us free of this difficulty.
7.2.2 Technical Aside: Convergence of Markov Processes . This section included for
completeness only. Not for a development course.
Let P be a transition probability on some state space X. It will be useful to have notation
for the m-step transition probability generated by P. This is, intuitively, the probability of
the system being in the subset A after m periods, starting from some given state x “today”.
Clearly, this is given by the measure µm , starting from the case where µ0 assigns probability
one to x. This measure we will denote in transition probability form as Pm (x, A), for any
measurable subset A of X.
The fundamental condition to be investigated is
Condition M (Stokey and Lucas [1989]) There exist > 0 and an integer M ≥ 1 such that for
any event A, either (i) PM (x, A) ≥ for all x ∈ X, or (ii) PM (x, AC ) ≥ for all x ∈ X.
To appreciate condition M, let’s look at a case when it is not satisfied. Consider the familiar
two-state Markov chain in which πi j = 1 if and only if i , j, for i, j = 1, 2. Pick A = {1}. Then
for any positive integer M, PM (x, A) = 1 either if M is even and x = 1, or if M is odd and
x = 2. Otherwise, PM (x, A) = 0. This means that condition M fails. We see therefore, that
the real bite of condition M is in the postulated uniformity with which all states hit particular
events.
The main result of this section is

Proposition 7.5. Under condition M, there exists a unique invariant probability measure µ∗ such
that for any initial µ0 on X, the generated sequence {µt } converges strongly to µ∗ .
Proof. We will follow the finite horizon case exactly. That is, we will show that
(1) M — the set of all probability measures on X — equipped with the total variation
metric is a complete metric space.
(2) The operator TM : M → M given by
Z
M
T (µ)(A) ≡ PM (x, A)µ(dx),
X
for all A, is a contraction.
(3) Thus TM has a unique fixed point µ∗ and the M-step iterates of any initial probability
measure must converge to µ∗ .
(4) The convergence of the entire sequence of measures, and not just this particular
subsequence, can then be established by a subsequence argument identical to that
used in the finite horizon case.
All the new stuff is in the first two items. To these we now proceed.
First we establish the completeness of M. To this end, suppose that {µn } is a Cauchy sequence
in M. then from the definition of the total variation metric, it follows that for each event A,
µn (A) is a Cauchy sequence of numbers. By the completeness of the real line, µn (A) converges
to some µ(A) for each A. We will show that µ is a probability measure and that µn converges
strongly to µ.
It is obvious that µ(A) ∈ [0, 1] for all A, that µ(X) = 1, and that µ(∅) = 0. It remains to prove
countable additivity to establish that µ is indeed a probability measure. To this end, let {Ai }
be a countable collection of disjoint events in X. Then
X∞ ∞
X
µ(∪∞ A
i=1 i ) = lim µn ∞
(∪ A
i=1 i ) = lim µ n
(Ai ) = lim µn (Ai ) = sum∞
i=1 µ(Ai )
n→∞ n→∞ n→∞
i=1 i=1
where the second-last equality follows from the dominated convergence theorem. This
proves that µ is indeed a bonafide probability measure. What’s left to do is to show that
||µn , µ|| → 0 as n → ∞. Note that because {µn } is Cauchy, for all > 0, there is N such that if
n, m ≥ N,
|µn (A) − µm (A)| ≤
for all sets A. Taking limits in m, it follows that
|µn (A) − µ(A)| ≤
for all n ≥ N, and for all A, which implies that ||µn , µ|| → 0 as n → ∞.
This establishes the completeness of M.
Our next task is to show that TM is a contraction of modulus 1 − . To this end, pick µ
and µ0 in M. Then there is a “common” part γ and “idiosyncratic” parts µ1 and µ2 , so that
µ = µ1 + γ, µ0 = µ2 + γ, and µ1 and µ2 have disjoint support.
Digression. The technical details of this assertion rely on the Radon-Nikodym Theorem (see
Stokey and Lucas [1989, Lemma 7.12]). But one can informally illustrate how the common
part is obtained in the case where µ and µ0 have densities f and f 0 on the real line. In this
case, simply define g(x) ≡ min{ f (x), f 0 (x)} for each x ∈ R, and integrate this as you would to
get a cdf, to arrive at the common part of the measure γ. This measure γ is, of course, not
a probability measure. The idiosyncratic residuals µ1 and µ2 are then defined by inserting
the rest of the probability, event by event, to bring total probability up to µ and µ0 . In this
way you see that the measures µ1 and µ2 are not probability measures either. Now study
the values µ1 (X) and µ2 (X), which are the values assumed by the measures on the entire
state space. Note, first, that these must be equal to each other. Note, also, that their common
value must be equal to the supremum difference between the probability measures µ and µ0
over any event. To see this, note that for any event A
|µ(A) − µ0 (A)| = |µ1 (A) − µ2 (A)|

≤ max{µ1 (A), µ2 (A)}
≤ max{µ1 (X), µ2 (X)}.
This ends our digression.
Returning to the main argument, we see that

Z Z
||T µ, T µ || = sup | P (x, A)µ(dx) −
M M 0 M
PM (x, A)µ0 (dx)|
A
Z Z
= sup | P (x, A)µ (dx) −
M 1
PM (x, A)µ2 (dx)|.
A
Now consider any event A and its complement AC . Without loss of generality suppose that
PM (x, A) ≥ for all x ∈ X. If K denotes the common value of µ1 (X) and µ2 (X) (see digression
above), then it must be the case that
Z Z
M 1
| P (x, A)µ (dx) − PM (x, A)µ2 (dx)| ≤ (1 − )K.
Combining these last two observations, and the observation in the digression, we see that
||TM µ, TM µ0 || ≤ (1 − )||µ, µ0 ||,
which completes the proof that TM is a contraction.
7.3 Interactive Inequality
It is now time to unpack the nature of the household production function. In many important
situations, it is determined by relative prices. This nring us to models of interactive
inequality, in which no household or dynasty can be treated as an isolated unit.
7.3.1 Relative Prices and Efficiency Units. Economists have traditionally employed a
simple shorthand for the study of occupational diversity, which is to reduce different
qualifications and skills to aggregate quantities of “human capital”. In other words, all
human capital is — even before we write down the definition of equilibrium for the society in question
Returns
“Financial Capital”
“Human Capital”
Investment Levels
Figure 7.3. Efficiency Units
— commonly expressible in some common efficiency unit.2 This approach is summarized

by Becker and Tomes in their 1986 paper:
“Although human capital takes many forms, including skills and abilities,
personality, appearance, reputation and appropriate credentials, we further
simplify by assuming that it is homogeneous and the same “stuff” in
different families.” (Becker-Tomes (1986, p.56), emphasis ours)
The crucial assumption is that the relative returns to different occupations are exogenous,
so that the reduction to efficiency units can be carried out separately from the behavioral
decisions made in the population. It is, in fact, common to specify that the returns to human
capital are concave while the return to financial investment is linear (see Figure 7.3).
It is unclear what different “levels” of human capital mean independent of the relative market
returns, which are typically endogenous. We might all agree that skilled labor embodies
more human capital than unskilled labor. But the assumption that skilled and unskilled
can be reduced to a common and determinate yardstick of efficiency units presumes much
more than this ordinal comparison. The implications of such an assumption can be quite
significant, as we show in the next section.
2So, for instance the investment choice in Loury’s model is interpreted as a choice of “how much” education to
acquire: there is no formal difference between human and physical capital. The so-called endogenous growth
models (see, e.g., Lucas (1988)) continue, by and large, to retain this shorthand.
7.3.2 Endogenous Relative Prices: Persistent Inequality With Two Skills. The follow-
ing analysis is based on Ray (1990). Elements of this model appear in Ljungqvist (1993),
Freeman (1996) and Mookherjee and Ray (2003). The specific exposition follows Ray (2006).
Suppose that aggregate production is a CRS function of just two inputs: skilled and unskilled
tasks, satisfying Inada conditions in each. There are two occupations: skilled and unskilled
labor. The latter can only do the unskilled tasks. Wages in each occupation equal their
respective marginal products, and so depend on relative supplies of workers in the two
occupations. Skill acquisition requires a fixed parental investment in education. Assume
that this is the only way a parent can transfer wealth to their children, i.e., there are no
financial bequests. In a later section we drop this assumption.
Now observe that even if all parents in the economy have identical wealth and preferences,
they cannot all leave the same bequests. The reason is simple. If every parent keeps their child
unskilled, there will be no skilled people in the next generation, raising the return to skilled
labor enough that investment in skill will be the optimal response. Conversely, if all children
are skilled, the return to skill will vanish, killing off the investment motive.3 Hence even
if all families start equal in generation 0, some will invest and others will not; in the next
generation their fortunes must separate.
Just who goes in one direction and who in another is entirely accidental, but such accidents
will cast long shadows on dynastic welfare. Indeed, the two directions are utility-equivalent
for generation 0, but not for generation 1! Furthermore, in succeeding generations wealthier
parents will have a greater incentive to train their children, so that the “primitive inequality”
that sets in at the first generation will be reinforced: children of skilled parents will be
more likely to acquire skills themselves. The logic of “symmetry-breaking” implies that
every steady state in this example must involve persistent inequality. The endogeneity of
occupational returns is central to this argument.4
In contrast to this example, observe that the same argument does not apply to activities in
which each unit is a perfect substitute for another. For instance, if shares in physical capital
can be divisibly held, everyone can derive the very same rate of return on each unit. But a
single individual cannot hold an arbitrarily fine portfolio of different occupations.5
In the subsections that follow, I develop this argument in more detail.
7.3.2.1 Preliminaries Time is discrete, running t = 0, 1, 2 . . .. A dynasty is represented by

an infinite sequence of individuals, each individual living for a single period. There is a
3If skilled labor cannot perform unskilled tasks, the unskilled wage will become very high by the Inada
conditions. But even if they can perform unskilled tasks, this will equalize the two wage rates. Either
interpretation has the same outcome.
4See e.g., Katz and Murphy (1992) for the responsiveness of US skill premia to relative supply of skilled workers.
5This discussion suggests, then, that the correct dividing line is not between “physical” and “human” bequests,
but rather bequests that result in endowments that are alienable (e.g. money) and endowments that are not
(e.g. occupations). The latter may include transfers of physical assets such as a family business which is not
incorporated — perhaps for reasons of moral hazard or simply the lack of development of a stock market. These
transfers are no different from human bequests in their implications for disequalization, and should be included
in the category of “occupational bequests”.
continuum of dynasties so that a unit mass of atomless individuals belongs to a generation

at each date.
There are two skill categories, “high” and “low”, which are combined via a production
function f to produce a single final output, which we take to be the numeraire. An individual
in the high-skill category (or a “high individual” for short) earns a wage w̄t at date t.
Likewise, a low individual earns wt . Whether or not an individual is high or low depends
¯
on the investment made by her parent. Being low at any date requires no investment by the
parent; being high requires an exogenous investment of x.
Earned income is partly consumed and partly used in educating the individual’s offspring.
Depending on the education level of the child, the child receives an income next period, and
the entire process repeats itself without end.
Now I turn to the determination of wages. Assume that the production function f for final
output is smooth, CRS in its two inputs, strictly concave in each input and satisfies the Inada
end-point conditions. Given a unit mass of individuals, if a fraction λ of them is high at
some date, then the high wage is given by
w̄(λ) ≡ f1 (λ, 1 − λ),
while the low wage is given by
w(λ) ≡ f2 (λ, 1 − λ).
¯
where these subscripts represent partial derivatives. We will call these wages the wages
associated with λ.
It is easy to see that w̄(λ) is decreasing and continuous in λ, with w̄(λ) → ∞ as λ →
0. Likewise, w(λ) is increasing and continuous in λ, with w(λ) → ∞ as λ → 1. These
¯ ¯
observations imply, in particular, that there exists a threshold λ̃ such that w̄(λ̃) = w(λ̃).
¯
To complete the description of the model, we presume that each generation t maximizes an
additive function of the one-period utility ut from its own consumption, and the lifetime
utility (Vt+1 ) felt by generation t + 1, discounted by δ ∈ (0, 1). The utility function u will
be assumed to be increasing, smooth and strictly concave in consumption, and defined at
least on [−x, ∞). This last requirement is innocuous but serves to simplify notation and
exposition. Moreover, the idea that consumption can go negative captures the idea that the
borrowing constraint is never absolute, but that the investment of x at lower wealth levels
entails ever greater utility losses (by strict concavity of u).
7.3.2.2 Equilibrium Suppose, now, that an infinite sequence of wages is given, one for
each skill category. We may denote this by the path {w̄t , wt }∞ . With such a sequence given,
¯ t=0
consider the maximization problem of generation t. Denote by V̄ t the lifetime utility for
a high member of that generation, and by V t the corresponding lifetime utility for a low
member. Standard arguments tell us that the¯ sequence {V̄ t , V t }∞t=0
is connected over time in
the following way: for each date t, ¯
(7.6) V̄ t = max u(ct ) + δVt+1

subject to the conditions that
(7.7) ct + xt = w̄t ,
and
Vt+1 = V̄ t+1 if xt ≥ x
(7.8) = V t+1 if xt < x.
¯
In exactly the same way, V t = max u(ct ) + δVt+1 , subject to the analogous budget constraint
ct + xt = wt and (7.8). ¯
¯
These maximization problems describe how education levels change from generation to
generation, given some sequence of wage rates. To complete the equilibrium setting, we
remind ourselves that the wages are endogenous; in particular, they will depend on the
proportion of high individuals at each date.
Formally, for given λ0 ∈ (0, 1), a competitive equilibrium is a sequence {w̄t , wt , λt }∞
t=0
such that
¯
[i] Given λ0 , the path {λt } is generated by the maximization problems described above.
[ii] For each t, w̄t and wt are the wages associated with λt .
¯
Standard fixed-point arguments suffice to show that a competitive equilibrium exists, and
we will not pursue such matters here.
Notice that our definition of competitive equilibrium assigns wages on the presumption that
skilled labor must carry out skilled tasks. Alternatively, we could restate the definition so
that if the “natural wages” (as given by marginal product) for a skilled worker falls short of
that of an unskilled worker, the former will move into the sector of the latter so that the two
wages will be ex post equalized. None of this matters much anyway because of the following
easy observation, which holds no matter what definition we use:
Observation 7.1. Recalling that λ̃ solves w̄(λ̃) = w(λ̃), 0 < λt < λ̃ for all t ≥ 1 along any
competitive equilibrium. ¯
The formalities of the (obvious) proof are omitted.6 From now on I will also presume that
λ0 ∈ (0, λ̃) as well. There is no great mystery in this: it saves the expositional trouble of
having to qualify several arguments for the initial value of λ.
For later use, I also record a familiar single-crossing observation.
Observation 7.2. Under a competitive equilibrium, there is no date at which a low person creates a
high child while simultaneously, a high person creates a low child.
The proof of this observation runs exactly parallel to (and in fact can be derived as a special
case of) Proposition 7.1, and is therefore omitted.
6That λ > 0 for all t follows from the fact that the difference between skilled and unskilled wages would be
t
infinitely high otherwise, so that some educational investment would have taken place prior to that period.
[Here we use the assumption that u is defined on [−x, ∞).] On the other hand, λt cannot exceed λ̃ for any t ≥ 1,
for in that case high and low wages are equalized, and no one in the previous generation would then have
invested in high skills.
7.3.2.3 Steady States A fraction λ is called a steady state if there exists a competitive
equilibrium {w̄t , wt , λt }∞
t=0
from λ with (w̄t , wt , λt ) = (w̄, w, λ) for all t, where w̄ and w are
¯
the wages associated with λ. ¯ ¯ ¯
The single-crossing property in the previous section yields a simple characterization of

steady states. Let w̄ ≡ w̄(λ) and w ≡ w(λ) be the associated wages, and let V̄ and V be the
¯
lifetime utilities associated with being¯(initially) high and low, respectively. By Observation
¯
7.2, the following two conditions are necessary and sufficient for λ to be a steady state:
V̄ = u(w̄ − x) + δV̄ ≥ u(w̄) + δV,
¯
while
V = u(w) + δV ≥ u(w − x) + δV̄.
¯ ¯ ¯ ¯
Combining these two expressions, we may conclude that
u(w̄) − u(w̄ − x) ≤ δ(V̄ − V) ≤ u(w) − u(w − x)
¯ ¯ ¯
is a necessary and sufficient condition for λ to be a steady state. Combining this expression
for the values of V̄ and V, we have established
¯
Proposition 7.6. The fraction λ (with associated wages (w̄, w)) is a steady state if and only if
¯
δ
(7.9) u(w̄) − u(w̄ − x) ≤ [u(w̄ − x) − u(w)] ≤ u(w) − u(w − x)
1−δ ¯ ¯ ¯
κ_
_
κ
~ λ
λ3 λ2 λ1 λ∗ λ
Figure 7.4. The Set of Steady States

Figure 7.5 plots the three terms in (7.9) as a function of λ. The left hand side, denoted by κ̄,
is just the utility cost to a high parent of acquiring skills for her child. This lies uniformly
below the right hand side, denoted by κ, which tracks the same utility cost to a low parent.
Finally, the middle term, denoted by B,¯ is the present value of benefits to being high rather
than low. Of course, the κ̄ and κ lines meet at λ̃, because wages are equalized there.
¯
Note, moreover, that as λ → λ̃, B turns negative while κ̄ is positive. On the other hand,
as λ → 0, B grows unboundedly large while κ̄ is bounded above. Because the changes
are monotone, there is a unique λ∗ ∈ (0, λ̃) such that the first inequality in (7.9) holds with
equality. Observe, moreover, that at λ = λ∗ , the second inequality in (7.9) must hold as well,
because of the strict concavity of the utility function. Thus the set of steady states contains
some interval to the left of λ∗ , and must be a subset of (0, λ∗ ].
Beyond this last observation, the set of steady states may be complicated. In particular, the
set need not be connected. For instance, in Figure 7.5, the set of steady states is the union of
the two intervals [λ3 , λ2 ] and [λ1 , λ∗ ].
Notice that every steady state involves persistent inequality, not just in gross incomes but in
net utility. In the subsection that follows, I show that the dynamics of an intertemporal
competitive equilibrium leads society to one of these steady states.
7.3.2.4 Dynamics Recall from the previous section that λ∗ is a steady state, and it is the
largest possible steady state. We now state and prove the following
Proposition 7.7. If λ0 > λ∗ , then there exists a unique competitive equilibrium from λ0 . It goes to
a steady state in one period: λ0 > λ1 = λt for all t ≥ 1.
If λ0 < λ∗ , then along a competitive equilibrium λt converges monotonically to the smallest steady
state no less than λ. Convergence is never attained in finite time unless λ happens to be a steady
state to start with, in which case λt = λ0 for all subsequent t.
The proposition provides a full account of the behavior of skill proportions over time,
starting from any initial condition. If that initial condition happens to be a steady state,
the proposition rules out any equilibrium path other than the steady state path itself. More
interesting is the asymmetry of equilibrium behavior under the two remaining kinds of
initial conditions. When λ0 is larger than the largest conceivable steady state, convergence
to a steady state occurs in a single unit of time. When λ0 is such that there are steady states
“above” it, convergence is gradual in that the process is never completed in finite time. This
asymmetry may have interesting implications for unanticipated technical changes which,
once realized, are expected to stay in place thereafter. Changes that call for a reduction
in steady state skill proportions take place quickly and dramatically, whereas a climb to a
higher steady state is more gradual and drawn out.
Moreover, in the case that convergence is “up” to a steady state, the proposition asserts that
it will occur to the nearest steady state to the right of the initial conditions. Put another way,
only the left-most steady state in each interval of steady states can be an attractor for initial
conditions that are distinct from that steady state, and the basin of attraction is precisely the
set of initial conditions that lie between it and the next, lower interval of steady states (if
any). Thus, despite the multiplicity of steady states, final outcomes can be tagged to initial
conditions in a unique way, allowing us in principle to perform comparative dynamics.
To provide some intuition for these results, first study a non-steady-state value of λ that is
smaller than λ∗ . Why is this not a steady state? Surely, the “no-deviation” condition for the
skilled (the first inequality in (7.9)) is satisfied; after all, it was satisfied at λ∗ , and now at
the smaller value of λ, both the wage differential is higher and the utility cost of education is
lower for the skilled. So, the reason why λ fails to be a steady state is that the “no-deviation”
condition for the unskilled — the second inequality in (7.9) — is violated; their utility cost
of education is high, but not as high as the wage differential. To maintain equilibrium
incentives, then, the economy-wide skill ratio must rise, compressing the sequence of wage
differentials until the unskilled are exactly indifferent between acquiring and not acquiring
skills. [We have already seen in Lemma 7.1 that this compression to indifference is a necessary
feature of non-steady-state equilibrium, otherwise no one would stay unskilled.]
Now here is the main point: the skill ratio achieved in the very next period cannot be a steady
state. For if it were, then the no-deviation condition of the unskilled must be satisfied here,
so that the new implied wage differential generates no incentive for them to acquire skills.
But the very same wage differential created indifference at date 0, when the utility cost of
acquiring education was higher for the unskilled! This is a contradiction. By an obvious
recursive argument, it follows that the upward movement in the skill ratio must be gradual
and perennial.
Exactly the opposite is true for non-steady-state values of λ that exceed λ∗ . For such values,
the “no-deviation” condition for the skilled surely fails; the wage differential is too small
relative to the utility cost of maintaining skills. So in equilibrium, λ falls. This fall along the
equilibrium path raises wage differentials so that the skilled are now indifferent between
maintaining and relinquishing skills.
We claim that the new skill ratio one period later must be a steady state. Suppose the claim is
false. Then the new skill ratio is not a steady state, and this can happen for one of two
reasons. First, the no-deviation condition for the unskilled fails — the wage differential at
date 1 is too attractive. In that case, we already know that the remaining sequence of wage
differentials (counting from date two on) must render the unskilled indifferent at date 1. But
this means that the sequence of wage differentials counting from date one on is still attractive
for the unskilled, and would have been a fortiori attractive relative to the skill-acquisition
cost for the unskilled at the higher skill ratio prevailing at date 0. But this, in turn, exceeds
the cost of skill maintenance for the skilled, which contradicts the fact that the equilibrium
creates indifference for the skilled at date 0 (see last sentence of preceding paragraph).
The second reason why the new skill ratio may fail to be a steady state is that the no-deviation
condition for the skilled fails again (just as it did at date 0). This means that the skill ratio
must fall even further in succeeding periods to create indifference for the skilled at date 1.
However, because the wage differential at date 1 is not attractive enough, this means that
the entire sequence of wage differentials, counting the one at date 1, would not have been
attractive enough for the very same skilled individuals, had they been located at date 0, but
with the utility costs they possess at date 1. This would be a fortiori true if we were to replace
the utility costs with the true utility costs of maintaining skills at date 0, which are higher.
But now we have a contradiction again (see the last sentence two paragraphs above). This
completes our intuitive description.
Proposition 7.7 asserts convergence from every initial condition to some steady state.
Moreover, the particular steady state to which convergence occurs can be identified, and
this is of interest because there are multiple steady states.
How general are these results on convergence; in particular, do they extend to the case of
several occupations? The simple answer is that I do not know. At the same time, as we are
going to see in the next section, with endogenous relative prices the steady state findings are
very general. With endogenous price determination and imperfect credit markets, steady-
state inequality among ex-ante identical agents is inevitable under broad conditions. It
would, therefore, be of great interest to see if the dynamic counterparts of the results in this
section also carry over.
7.4 More on Relative Prices and Persistent Inequality
Recall the two-skill example with which we started the previous section. We argued that
even with parents of identical wealth, symmetry-breaking must occur “in equilibrium”:
after all, it cannot be that all parents skill their children, or that all parents leave their
children unskilled. This symmetry-breaking led to persistent inequality. It is natural to
ask whether the same results hold when parents can compensate low-skilled children with
greater financial bequests. Symmetry-breaking would still occur, but only in “occupational
space”: there would be no inequality in overall wealth, human plus financial. In this section,
based on Mookherjee and Ray (2007), I carefully investigate the validity of this argument.
A second assumption that we made is that there are only two occupations. On the face of
it, this is is just a simplification, and indeed, most of the literature works off two or three
occupations simply for expositional ease. But as we shall see, bringing in many occupations
(all with varying relative prices) makes two major conceptual differences. The first is that
“many” different levels of human capital can be chosen, along with financial bequests
(see previous paragraph). From the household point of view — with all relative prices
taken as given — we have then a standard “production function” of the form displayed
in Figure 7.3, which is the outer envelope of all human and financial investments. But
there is one difference: the shape of this “production function” is entirely dependent on
endogenous relative prices! Viewed from this perspective, at the household level all
returns to investments are driven by relative prices. Whether this gives rise to a concave
“technology” or not does not depend on technology but on economics, and on economic
equilibrium at that. Yet, we will have something quite definite to say about the curvature of
this production function.
The second conceptual difference, as we shall see below, is that steady state inequality might
remain but history-dependence tends to vanish! Recall in the the two-skill model (see Figure
7.5) that there isn’t just inequality in steady state, there is a continuum of steady states.
Indeed, the dynamics studied in that model take us to different limit points depending on
initial conditions. Thus the two-skill model displays not just persistent inequality, but also
persistent dependence on initial conditions, or on “history”.
It turns out that with a rich set of occupations, this history-dependence vanishes. This
doesn’t happen at the level of an individual dynasty, whose fortunes remain profoundly
linked to the distant past. Rather, it happens to the economy as a whole.
7.4.1 The Model. Occupations and Training. There is a compact measurable space H of
occupations that will be used in the production of a single, aggregative final good. There is
an exogenous training cost x(h) for occupation h ∈ H, denominated in units of final output.7
Production. A single aggregate output is produced by physical capital8 and individuals
who hold occupations in H.
Output y (net of the undepreciated capital stock) is produced by a continuous,9 strictly
quasiconcave CRS production function y = f (k, λ), where k is physical capital and λ is
an occupational distribution (a finite measure on H). It is helpful to interpret different
occupations as corresponding to different kinds of human capital.
Prices and Firms. Firms maximize profits at given prices. Normalize the price of final output
to 1. Let w ≡ {w(h)} denote the wage function, and p ≡ (r, w) the factor price function, where
r is the rate of interest. Denote by c(p) the unit cost function.
By constant returns to scale, profit maximization at positive output is possible if and only if
c(p) = 1; in that case call p a supporting price. If (k, λ) is a profit-maximizing choice under the
supporting price p, we will refer to it as an associated input vector.
In these notes, I assume that the rate of interest r is exogenously given and time-stationary.
One simple interpretation is that capital is internationally mobile and that our economy is
a price taker on the world market. Under this interpretation we also assume, in effect, that
people are not internationally mobile: the wage function w will be determined domestically.
See Mookherjee and Ray (2006) for extensions.
Now that r may be treated as a parameter, say that w is a supporting wage if p = (r, w) is a
supporting price.
Families. There is a continuum of families indexed by i ∈ [0, 1]. All families are ex
ante identical, so we endow [0, 1] with Lebesgue measure ν. Each family i has a single
representative at each date or generation, indexed by t. Call this agent (t, i).
Consider a member of generation t. She begins adult life with a financial bequest b and
an occupation h, both “selected” by her parent. The latter is obviously shorthand for the
assumption that the parent bears the costs of upbringing and education (the child can select
the particular occupation with no difference to the formal analysis). The overall wealth
of our generation-t adult is then W ≡ b(1 + r) + wt (h), where wt (h) is the going wage for
occupation h at date t.
7As in Mookherjee and Ray (2003), this may be generalized to allow training costs to depend on the pattern of
wages. We conjecture that the principal qualitative results of this paper will continue to hold in that setup.
8As long as capital goods are alienable and shares in them can be divisibly held, having several capital goods
makes no difference to the analysis.
9Endow the space of all nonnegative finite measures on H with the topology of weak convergence. We ask that
output be continuous with respect to the product of this topology and the usual topology on k.
The agent correctly anticipates factor prices pt+1 ≡ (r, wt+1 ) for the next generation t + 1, and
selects her own financial and educational bequests (b0 , h0 ) to maximize
(7.10) U (W − x(h) − b0 ) + V ((1 + r)b0 + wt+1 (h0 ))
subject to the no-intergenerational-debt constraint b0 ≥ 0. We assume that U and V are
smooth, increasing and strictly concave, and that U has unbounded steepness at 0.
Now b0 and h0 become the financial and educational inheritance of her child — generation
t + 1 — and the entire process repeats itself ad infinitum.
Clearly, we are using a specification of utility that allows for income-based altruism. More
on this specification later.
The condition b0 ≥ 0 is a fundamental restriction stating that children cannot be held
responsible for debts incurred by their parents. The capital market is active in all other
senses: households can make financial bequests at the going rate r, and firms can freely hire
in capital at the very same rate.
Equilibrium. Begin with an initial distribution of financial wealth and occupational choices.
A competitive equilibrium given these initial conditions is a sequence of wage functions wt , t =
0, 1, 2, . . . and occupational distributions λt , t = 0, 1, 2, . . ., as well as occupational and bequest
choices for each generation in each family — {ht (i), bt (i)} — such that for each t and each family
i:
(a) person (t, i) chooses (bt+1 (i), ht+1 (i)) to maximize the utility function in (7.10), given that
her own starting wealth equals (1 + r)bt (i) + wt (ht (i));
(b) these decisions aggregate to λt at each t:

λt (H) = ν{i ∈ [0, 1]|ht (i) ∈ H}
for every Borel subset H of occupations, and
(c) wt is a supporting wage, with associated input vector (λt , kt ) for some choice of kt .
Observe that equilibrium conditions place no restrictions on kt . Because there is international

capital mobility, financial holdings by households need bear no relation to capital used in
production.10
A steady state is a competitive equilibrium with stationary prices and distributions; (wt , λt ) =
(w, λ) for all t, and strictly positive output.
The following observation is useful: in steady state, the total wealth of every family, not just
of the economy as a whole, must be stationary.11
Observation 7.3. The wealth of every family is stationary in any steady state.
10When there is no international capital mobility, k must equal the aggregate of financial holdings, and r must
t
adjust to assure this equalization in equilibrium.
11In general, such an assertion is not true of the financial wealth or the occupational choice of a family, which
may vary over time.
Proof. This follows directly from Proposition 7.2.
7.4.2 The Becker-Tomes Benchmark. A special case of this model is the following
elementary textbook exercise: only financial bequests are possible (earning interest r), and
everyone earns a fixed wage w.
Stochastic shocks apart, this is exactly the specification for the Becker-Tomes (1979) model,
which assumes a linear rate of return to parental investment in children.12 We will henceforth
refer to this special case as the Becker-Tomes benchmark.
In this special case, a parent with wealth W simply selects b ≥ 0 so as to maximize
U (W − b) + V (w + (1 + r)b) .
Let the resulting wealth of the child be denoted W̃ ≡ w + (1 + r)b. We may write W̃ as a
function of W, w and r: W̃(W; w, r). By our assumptions, W̃ is fully characterized by the first
order conditions
!
W̃ − w
(7.11) 0
U W− ≥ (1 + r)V 0 (W̃),
1+r
with equality if W̃ > w.
It is obvious that W̃(W; w, r) is nondecreasing and continuous in W. So an iteration of this
mapping from any initial condition W > 0 will yield long-run wealth starting from W. Call
this long-run wealth Ω. Passing to the limit in (9.42), it is trivial to see that if w ≤ Ω < ∞,
0 rΩ + w

(7.12) U ≥ (1 + r)V 0 (Ω), with equality if Ω > w.
1+r
Becker and Tomes impose the following restriction on bequest behavior: ∂∂W W̃
∈ (0, 1),
justifying it by available empirical evidence. This implies that the wealth of all families
will converge to a common limit Ω, independent of initial wealth. To ensure that our larger
model is consistent with this key equalization property of financial bequests, we impose a
similar (though weaker) restriction in this special case:
[LP] Limited Persistence. For any r > −1 and w ≥ 0, there is at most one solution in Ω ≥ w to
(7.12).
As in Becker-Tomes, [LP] also implies that limit wealth Ω(w, r) is well defined and
independent of starting wealth as long as as that starting wealth strictly exceeds w (that
is, as long as starting financial wealth is positive). Given [LP], here is how we define that
12A more sophisticated version of their model is Becker and Tomes (1986), in which bequests in the form of
human capital are also permitted, but human capital is a priori reduced to efficiency units and it is assumed that
the rate of return to successive units of human capital is declining. In that variant, all families will aim to invest
the (same) amount of human capital before turning to linear financial bequests. Indeed, we do not claim that our
elementary textbook exercise captures the full implications of the Becker-Tomes models. An important theme
in these papers is the interplay between luck and convergence, an issue that is not of relevance here.
limit wealth: Set Ω(w, r) equal to Ω, where Ω solves (7.12), in all cases except the one in
which
rΩ + w

U0 > (1 + r)V 0 (Ω)
1+r
for all Ω > w, in which case set Ω(w, r) = ∞.13 The important point is that Ω(w, r) is the same
no matter what the (positive) level of inital wealth is.
This condition is less restrictive than the original Becker-Tomes assumption on ∂W

0
∂W
, though
the reader is welcome to keep the stronger restriction in mind (as long as it is understood
that an implicit restriction on the interest rate is also implied thereby).
Condition [LP] is extremely easy to check. The following observation illustrates this by
studying the HARA class of preferences.
Observation 7.4. Suppose preferences satisfy the following restriction: There exists δ > 0 such that
V = δU, and U belongs to the HARA family:
(H) − U00 (c)/U0 (c) = 1/(α + βc),
where (α, β) ≥ 0 and nonzero. Then the limited persistence property is satisfied, with the single
exception in which α = 0 and w = 0, in which case it is satisfied for all but one value of r.
Provided that preferences satisfy a constant discount rate property, Observation 7.4 states
that [LP] holds for utility functions that are iso-elastic or exponential, or belong to the HARA
class which nests these as special cases.
It helps to illustrate bequest behavior in the setting with financial bequests alone for the
special case of iso-elastic utility with discounting: U(c) = (c1−σ − 1)/(1 − σ) with σ > 0, and
V ≡ δU, with δ ∈ (0, 1). Define ρ ≡ [δ(1 + r)]1/σ . Intergenerational wealth movements in the
Becker-Tomes benchmark with stationary (w, r) then takes the form:
(1 + r)ρ ρ
W̃ = W+ w
1+ρ+r 1+ρ+r
if W ≥ w
ρ, and W̃ = w otherwise. This allows us to calculate limit wealth:
w if ρ ≤ 1,


ρ

Ω(w, r) =  ρ ∈ (1, 1 + 1r ),

(7.13)

 1−r(ρ−1) w if
if ρ ≥ 1 + 1 .

 ∞
r
If ρ ≤ 1, there are no (limiting) financial bequests in steady state in the Becker-Tomes
benchmark, and Ω(r, w) = w. Limit wealth Ω(w, r) is finite only if ρ ∈ (1, 1 + 1r ).
If r is high enough that this condition is not satisfied, limit wealth is infinite: our version
of the limited persistence property is satisfied,14 whereas the Becker-Tomes version is not.
We only point this out to emphasize that the Becker-Tomes version of limited persistence
13The only subtlety here is one in which (7.12) holds with equality at Ω = w, while the inequality > holds for
all Ω > w. In this case both limits w and ∞ are potential candidates, but the correct limit for starting W > w is
easily seen to be the latter. This is a nongeneric case of little import but in any case our definition handles it.
14Notice that there are two limit wealths in this case, one at infinity and one at zero. But there is only one limit
wealth provided we start with strictly positive wealth, and this is what [LP] requires.
imposes more than a restriction on preferences, but in any case we are also fundamentally
interested in the case in which limit wealth is finite.15
A Comment on Nonpaternalism. We end this section by pointing out that preferences
are inconsistent with the limited persistence assumption. By the envelope theorem, the
derivative of the value function is just the marginal utility of (equilibrium) consumption.
At a steady state, then, these marginal utilities drop out from the Euler equation, and the
conditions become independent of wealth. In particular, an increase in parental steady
state wealth translates into an equal increase in child wealth, so that [LP] fails. Dynastic
preferences do not allow us to take the Becker-Tomes postulates fully on board.
Why do we want to incorporate [LP] in the first place? One answer is that it is empirically
attractive. That may be so, but we do not impose [LP] for this reason. We do so because
we want to show that this assumption generates a form of the steady state wage function
that can never exhibit diminishing returns! If at all that wage function is nonlinear, it must
be nonconvex at least over a region. This shows that the assumption that returns to human
capital must be diminishing (Becker and Tomes (1986)) merits careful scrutiny, to say the
least. None of these points emerge with dynastic preferences, and this — apart from financial
bequests — is another basic difference from the analysis in Mookherjee and Ray (2003).
7.4.3 Two Occupations with No Financial Bequests. In what follows, we study an

extremely simply case, one in which there are just two occupations, and financial bequests
are nonexistent. You can suppose that we are in a world in which Ω(w, r) = w for all w.
Call the two occupations “skilled” and “unskilled” labor. For unskilled labor take the
training cost to be zero. For skilled labor assume that there is a exogenous training cost X,
which is just the number of units of the consumption good used as input into the training
process. Let λ denote the fraction of the population at any date that is skilled. If some
well-behaved production function f (satisfying the usual curvature and Inada end-point
conditions) determines the wage to skill categories, the skilled wage at that date will be
given by w̄(λ) ≡ f1 (λ, 1 − λ), while the unskilled wage will be given by w(λ) ≡ f2 (λ, 1 − λ).
¯
where subscripts denote appropriate partial derivatives.16 This yields the following simple
characterization: a fraction λ of skilled people is compatible with a steady state if and only
if
U (w̄(λ)) − u (w̄(λ) − X) ≤ V (w̄(λ) − x) − V (w(λ))]
¯
(7.14) ≤ U (w(λ)) − U (w(λ) − X)
¯ ¯
The left hand side of (7.14) represents the utility sacrifice of a skilled parent (hereafter
15While the case of unbounded limiting wealth is formally a special case, ongoing growth really calls for a
different model, ideally one in which training costs are endogenous as well.
16This applies only in the unrealistic event that skilled workers cannot perform unskilled tasks. More generally,
if skilled workers can perform unskilled tasks, then the skilled wage cannot ever fall below the unskilled wage.
So when the skill intensity λ is large enough that f1 < f2 , wages will not be given by f1 and f2 , but will be
equalized (as a result of skilled workers filling unskilled positions whenever the latter pay higher wages). We
omit this minor complication here because a competitive equilibrium with a positive fraction of skilled workers
will never give rise to wage differentials that are incompatible with incentives for parents to educate their
children.
κu(λ)
b(λ)
κs(λ)
λ
λ6 λ5 λ4 λ3 λ2 λ1
Figure 7.5. Education Costs and Benefits in Two-Profession Model
denoted by κs (λ)) in educating its child, while the right hand side is the corresponding
sacrifice for an unskilled parent (denoted by κu (λ)). The term in the middle is the present
value benefit of all successive descendants being skilled rather than unskilled (which we
shall denote by b(λ)).
These benefit and sacrifice functions are illustrated in Figure 7.5. λ1 ∈ (0, 1) denotes the skill
intensity of the population at which the skill premium just disappears and the wages of the
skilled and unskilled are equal. So κs and κu intersect there. Likewise, λ2 is the point at
which the wages of the skilled net of training equal those of the unskilled. So b drops to zero
there. These observations can be used in conjunction with (7.14) to establish:
Proposition 7.8. There is a continuum of steady states in the two-profession model with exogenous
training costs, and both per capita income and consumption rise as the skill proportion in steady state
increases.
Proposition 7.8 tells us that multiplicity — in the sense of a continuum of steady states — is
endemic for a small number of professions. While stated only for the two-profession case, it
is easy enough to extend the argument to any finite number of distinct professions.
Notice that the structure of the set of steady states may be complicated. In particular, the set
need not be connected. For instance, in Figure 7.5, the set of steady states is the union of the
two intervals (λ6 , λ5 ) and (λ4 , λ3 ).
The proposition also states that steady states are ordered not only in terms of skill premium
but also per capita income: a steady state with a higher λ and lower skill premium
corresponds to higher per capita income net of training costs. This does not, however,
imply that these steady states are Pareto-ordered. For a detailed discussion of efficiency, see
Mookherjee and Ray (2003).
7.4.4 A Rich Set Of Occupations. For a while, now, we shall abandon the idea of just two
occupations and go to the opposite extreme, in which we examine what happens with a rich
set of occupations. Two assumptions characterize richness:
[R.1] The set of all possible training costs is a compact interval of the form [0, X].
[R.2] For every subset C ⊆ [0, X] of positive Lebesgue measure, if the occupational
distribution has zero value over every occupation h with x(h) ∈ C, then no output can
be produced.
Thus we don’t ask that every occupation be essential, only that (almost) every training cost
in [0, X] has an essential occupation attached to it. Observe that conditions [R.1] and [R.2]
really go together as a pair: without some restriction like [R.2], [R.1] can always be trivially
met by simply inventing useless occupations to fill up the gaps in training costs.
Together, [R.1] and [R.2] imply that whenever positive output is produced, the inhabited
range of “equilibrium training costs” is always equal to [0, X].
We justify richness by noting that while there are large differences in training costs between
unskilled occupations (such as farm workers or manual jobs) and skilled occupations (such as
engineers, doctors and lawyers), there are also many semi-skilled occupations (technicians,
nurses and clerks) with intermediate training costs and wages. Besides, there are large
differences in the quality of education within any given occupation, which translate into
corresponding differences in education costs and wages.
The methodological innovation in the richness asumption is that it allows families to fully
fine-tune their investments. Whether or not their investment set is convex depends, then,
not on assumed indivisibilities in training costs but in the endogenously determined factor
price schedule.
Note that [R.1] includes a bit more than occupational richness. It states that there is an
occupation with zero training cost, a restriction that we impose for expositional ease.
7.4.5 Steady States With Rich Occupations. Fix a steady state. Say that an occupation
(or training cost) is inhabited if some family chooses that occupation (or incurs that training
cost). By the richness conditions, we know that every steady state (which has positive output,
by definition) must exhibit a full measure of inhabited training costs. Suppose that we can
alter the wage function on the small set of uninhabited occupations without changing any of
the observed features of the steady state. Then we will say that the new steady state wage
function (which only differs by specifying different wages for uninhabited occupations) is
an equivalent representation of the old.
One particular representation is of interest, in which all occupations with the same training
cost command the same wage. In that case (though with some abuse of notation), we shall
go back and forth between w(h) and the representation w(x).
Proposition 7.9. Assume [R.1], [R.2] and [LP].
(a) Every steady state has an equivalent representation with a continuous wage function.
(b) In an equal steady state, this equivalent representation may be described as follows: there exists
w ≥ 0 such that
(7.15) w(x) = w + (1 + r)x
for all x. In that steady state, all families attain a common wealth of Ω(w, r), and
Ω(w, r) − w
(7.16) X≤ ,
1+r
where X is the highest training cost across all occupations.
(b) In an unequal steady state, the equivalent representation may be described as a two-phase wage
function: there exists w ≥ 0 and θ ∈ [0, X) such that for all occupations h with x(h) ≤ θ, (7.15)
holds:
w(h) = w + (1 + r)x(h).
Families that choose any of these occupations at any date all attain a common wealth of Ω(w, r) that
is precisely equal to w(θ).
On the other hand, for all occupations h with x(h) > θ, w and x are connected via the following
differential equation:
U0 (w(x) − x)
(7.17) w0 (x) = .
V 0 (w(x))
with endpoint constraint that the wage at cost θ equals w(θ) = w + (1 + r)θ.
Families in such occupations attain a wealth that is strictly greater than Ω(w, r), and the marginal
rate of return to these occupations, w0 (x), strictly exceeds 1 + r almost everywhere.
The outline of the argument is as follows. Since (almost) every training cost must have
at least one inhabited occupation, the marginal rate of return on training costs must be at
least r everywhere. In other words, the presence of human capital allows parents to transfer
wealth to their children at (weakly) lower cost than in a Becker-Tomes benchmark where
only financial bequests are possible. Hence every family must attain at least the wealth
Ω(w, r) that would have arisen in the latter context (corresponding to the flow wage w that
is available to every generation even in the absence of any educational investment). In an
equal steady state, the rate of return on human capital at all levels is exactly r, so that the set
of investment options is exactly the same as in a Becker-Tomes benchmark with stationary
(w, r). Therefore all families attain precisely the wealth Ω(w, r).
Finally, equation (7.16) must hold in such a steady state, because the occupation with the
highest training cost X must have a wage of w + (1 + r)X, and must be willingly chosen by
some family.
In an unequal steady state it is trivial to see that the rate of return to some occupation must
then exceed r. If not, the set of investment options would be just as in a Becker-Tomes
benchmark with financial bequests alone, and wealth equality is then the only possible
long-run outcome.
What is more subtle is the exact form the wage function must take. The proposition claims
that the wage function has two phases. For low-end occupations, financial and human
rates of return coincide and the wage function is linear. This phase has an endogenous
delineation: any occupation with training cost below the steady state financial bequest in
the Becker-Tomes benchmark with (w, r) must generate a rate of return of exactly r. (Note
well that w, the lowest wage, is endogenous.)
For occupations with training costs that exceed the Becker-Tomes steady state bequest, the
rates of return must be high enough to induce willing settlement. This requirement creates
at least a local nonconvexity of the steady state wage function: the marginal rates of return
to such occupations will exceed r.
A closer inspection of the differential equation (7.17) reveals that the shape of the wage
function in the second phase relies entirely on preferences. To be sure (and as we shall see
more clearly below), the existence and range of this phase will depend, among other things,
on the technology. For a large class of preferences, the wage function exhibits a “global”
nonconvexity, in the sense that the marginal rate of return rises monotonically with training
costs beyond θ.
Observation 7.5. Consider either the constant elasticity case with U(c) = (c1−σ − 1)/(1 − σ), σ > 0,
or the case of exponential utility U(c) = − exp(−αc), α > 0, and V ≡ δU. Then the marginal rate of
return on occupations monotonically increases with training cost beyond the boundary θ described
in Proposition 7.9.
In the case of exponential utility, the wage function takes the form (for x > θ):
1 1
(7.18) w(x) = exp(αx) + Ω(w, r) − exp(αθ).
δα δα
In the constant elasticity case w0 (x) monotonically increases to a finite asymptote that strictly exceeds
1 + r.
This description stands the traditional theory on its head. That theory presumes — usually
by assumption — that the rates of return to human capital must be declining in training
cost (see, for instance, Loury (1981) and Becker and Tomes (1986)). Therefore the poorer
families make all the human capital investment, and once families are rich enough so that
the marginal return on human capital falls to the constant rate assumed for financial capital,
all other bequests are financial.
In contrast, Proposition 7.9 is stated in a context in which the relative earnings of different
occupations are allowed depend on the occupational distribution, for which there is
considerable empirical evidence (e.g., Katz and Murphy (1992)). The proposition then asserts
that the theory endogenously generates rates of return that run counter to the assumptions
made in the literature. Financial bequests are made at the low end, while “occupational
bequests” carry a higher rate of return and are made by richer families. If a rich set of
occupations is essential, the concavity of returns to human capital is never an equilibrium
outcome in any unequal steady state.
Financial bequests at high rate
Financial bequests at low rate
Occupational bequests alone
X x
Figure 7.6. A Extension to Three Phases
As an aside, we note that we haven’t yet provided conditions for a steady state to be unequal,
but will soon proceed to an analysis of this question.
It is important to see that the predicted shape of steady state wage functions depend
fundamentally on the limited persistence assumption. If several occupations are essential in
steady state, then to sustain the “wealth-spreading” necessitated by those occupations and
to counteract convergence, wage functions must display a nonconvexity. A dynastic model,
such as the one studied in Mookherjee and Ray (2003), fails [LP] and cannot deliver this
property of wage functions in steady state.
Are our predictions counterfactual? We think not. First, it is well known from inequality
decomposition studies that earnings inequality accounts for most of overall income inequal-
ity. For instance, Fields (2004) summarizes observations from several studies, writing that
that “labor income inequality is as important or more important than all other income sources
combined in explaining total income inequality”.
Second, there is evidence that within the class of financial bequests, which are admittedly
large for rich families, intentional bequests are not important. For instance, Gokhale et
al (2001) argue that most financial bequests in the US economy are unintentional, the
result of premature death and imperfect annuitization. In the iso-elastic example, this
would correspond to the case with ρ below unity. Our theory then predicts that there are
no intentional bequests anywhere in the wealth distribution, so human capital differences
entirely account for all inequality, perhaps supplemented by unintended financial bequests
(which we do not formally model).17
Third, it is important to remember that by “occupations”, we mean not just human capital but
every productive activity that is inalienable. This includes human capital but it is certainly
not restricted to it. In particular, it is possible to view large financial bequests observed at
17For a model of unintended bequests arising from uncertain life span, see, e.g., Fuster (2000).
the top end of the distribution as a form of occupational investment by parents, in the form
of transfer of ownership or control of (partly inalienable) business activities.
Finally, it is easy to generate variants of this model which exhibit financial bequests at both
upper and lower ends of the wealth distribution. For instance, suppose that financial wealth
can be left at two rates of return: one at the existing rate r and another at a higher rate r0 , only
accessible when the level of financial bequests crosses some threshold. Then it can be shows
(I do not do this here) that in general, a steady state wage function will have a “three-phase
property”. Phases I and II look just like they do in this proposition. In phase III there is a
return to linearity, this time at the higher rate r0 . Financial bequests are left at both ends of
the wealth distributrion. Figure 7.6 illustrates.
7.4.6 Existence and Uniqueness With Richness. Existence. Having gained an under-
standing of the structure of steady states in this model, we can now provide conditions on
the technology and preferences that guarantee the existence of a steady state.
Our definition of a steady state includes the requirement that output must be positive, so that
existence is typically nontrivial. Proposition 7.9 informs us that a steady state must assume
a particular form. In fact, it is easy to see that given some baseline wage w for unskilled labor,
that proposition fully pins down the wage function. The only scope for variation lies in w. It
comes as no surprise, then, that the existence of a (nondegenerate) steady state depends on
the economy being productive enough to sustain positive profit at one of these conceivable
wage functions.18
Of course, that isn’t enough. Proposition 7.9 describes what a steady state necessarily looks
like, but is silent on the question of whether such a description is indeed sufficient for all the
steady state conditions. This, too, will need to be addressed.
We proceed, then, by searching for a steady state using the features described in Proposition
7.9. To this end, we describe the family of all two-phase wage functions. Start with a given
baseline wage of w, and set w(x) = w + (1 + r)x for all training costs no greater than
Ω(w, r) − w
θ(w) ≡ min{ , X}.
1+r
(This corresponds to the old threshold θ used in Proposition 7.9, but now we make the
dependence on w explicit.) For occupations with higher training costs — if any — the wage
function is set to satisfy the differential equation (7.17):
U0 (w(x) − x)
w0 (x)) = ,
V 0 (w(x))
with the endpoint constraint that the wage for training cost θ(w) equals w + (1 + r)θ(w), or
equivalently, Ω(w, r).
18There is a family resemblance here to the existence of steady states in the multisectoral optimal growth model,
which requires a “productivity condition” (see Khan and Mitra (1986)). Fixed point arguments that neglect such
a productivity condition, as in Sutherland (1970), are bound to fail because they do not eliminate the trivial
outcome with zero output and wealth all around.
This procedure generates a unique wage function corresponding to any choice of w. Repeat
this procedure for every w ≥ 0: we now have the entire two-phase family, with either one of
the phases conceivably degenerate.19
Given what we know already, the following “productivity” condition is necessary for the
existence of a steady state with positive output:
[P] Unit costs c(r, w) are less than or equal to 1 for some two-phase wage function.
It is easy enough to rewrite [P] as a productivity condition. For instance, if the production
function is written as A f ( f, λ), where A is some Hicks-neutral productivity parameter, [P]
states simply that A is large enough.
As the following proposition reveals, condition P (in conjunction with the other maintained
assumptions) is also sufficient for the existence of a steady state with positive output.20
Proposition 7.10. Under [R.1], [R.2], and [LP], a steady state with positive output exists if and
only if [P] is satisfied.
Condition [P] isn’t at all difficult to verify, one way or the other. As an example, suppose
that each training cost x corresponds to a unique occupation (so name it x as well), and that
the production function takes the Cobb-Douglas form
Z X
ln y = (1 − α) ln k + α(x) ln(λ(x))dx + ln A,
0
R
where A is a productivity parameter, α(x) ≥ 0 and α(x)dx = α ∈ (0, 1). Then it is easy to see
that for any wage function w,
Z X
ln c(r, w) = (1 − α)[ln r − ln(1 − α)] + α(x)[ln(w(x)) − ln(α(x))]dx − ln A.
0
The
R verification of [P] therefore simply entails the choice of a wage function that minimizes
α(x)w(x), and then checking whether the resulting expression above is nonpositive.
Uniqueness. Proposition 7.9 already takes a significant step towards uniqueness by
establishing that any steady state wage function, or at least its continuous equivalent
representation, must lie in the two-phase class. As described in detail earlier, it is linear
with return r over a range of training costs, and then displays a marginal rate of return
that strictly exceeds r. In a broad class of cases (see, e.g. Observation 7.5), this marginal
return can be shown to be ever-increasing, though it will usually possess a finite asymptote.
Finally, the training-cost threshold separating the two phases precisely corresponds to the
Becker-Tomes limit bequest with wage equal to the lowest wage along this function.
19If θ(w) = X, the second phase is degenerate. Remember that at w = 0, Ω(0, r) is the limit of Ω(w, r) as w ↓ 0. It
may or may not equal 0, which is always trivially a limit wealth. If it is, the first phase is degenerate.
20If [P] fails, it is easy to construct a steady state with zero output: simply construct the two-phase wage function
that starts from a baseline wage of zero, and place all individuals in the zero-wage occupation.
To be sure, the same economic fundamentals are consistent — at least in principle — with
several wage functions drawn from this two-phase class. But this is what the uniqueness
proposition rules out.
Proposition 7.11. Assume [R.1], [R.2], and [LP]. Then, apart from equivalent representations
which change no observed outcome, there is at most one steady state.
As already discussed, this uniqueness proposition has far-ranging implications. Apart from
Mookherjee and Ray (2003), to be discussed in more detail below, this observation has gone
unnoticed in the literature because the literature typically concentrates on a sparse set of
occupations (usually two, as in Galor and Zeira (1993) or occasionally three, as in Banerjee
and Newman (1993)). In such cases multiplicity is indeed endemic, but once the set of
occupations expands such multiplicity must shrink. We reiterate that the expansion of the
set of occupations does not convexify the set of choices. Indeed, as Proposition 7.9 takes
pains to explain, equilibrium nonconvexity is the rule rather than the exception.
Following up on this point, our uniqueness proposition does not rule out the path-
dependence of economic fortunes for individual families. The identities of those who inhabit
the different occupational slots is up for grabs and may — will — depend on historical
accident. But their numbers cannot.
Proposition 7.11 is a substantial extension of the uniqueness theorem in Mookherjee and Ray
(2003) to a context in which financial capital co-exists with human capital. Indeed, given the
simplified context of our model,21 the uniqueness result of Mookherjee and Ray (2003) can
be seen very easily and intuitively.
Imagine reworking Proposition 7.9 by imposing the additional constraint that no financial
bequests are permitted. One would reasonably suppose, then, that the first phase of the two-
phase function would disappear, and that any steady state wage function must be governed
by the differential equation (7.17) throughout. Now it is easy to see why there can be only
one such wage function. If we begin at two different initial conditions and apply (7.17)
thereafter, the two wage trajectories cannot cross — a well-known property for this class
of differential equations. In short, if there are two steady state wage functions, one must lie
entirely above the other. But now we have a contradiction, for two wage functions ordered in
this way cannot both serve as bonafide supporting prices for profit maximization. We obtain
uniqueness when there are no financial bequests.
While this serves as some intuition for the result at hand, different considerations emerge
when financial bequests are permitted. Now crossings of the two putative steady state wage
functions cannot be ruled out by taking recourse to uniqueness theorems for differential
equations. After all, the behavior of the wage functions is not governed throughout by
(7.17); a nontrivial “first phase” makes an appearance. Instead, the formal proof must rely
on behavioral arguments, based on household optimization, to rule out such crossings.
21Apart from the central difference of financial bequests, there are two differences between our model and that
of Mookherjee and Ray (2003). First, they use a nonpaternalistic bequest motive and work with value functions.
Second, training costs are endogenously determined in their model. However, these differences are minor and
can be readily accommodated.
7.4.7 Conditions for Inequality. We are now in a position a key question: under what
conditions does the steady state of the model involve inequality rather than equality? In the
language of Proposition 7.9, when must the second phase of the two-phase wage function
necessarily be nonempty?
A simple preliminary exercise lays the groundwork for a complete characterization of this
question. This exercise concerns the production technology alone and has nothing to do
with preferences.
Consider the class of all linear wage functions of the form w(x) = w + (1 + r)x defined on all
of [0, X], parameterized by w ≥ 0.
Observation 7.6. Assume [P]. Then there is a unique value of w — call it a — and a corresponding
linear wage function w∗ with w∗ (x) = a + (1 + r)x for all x — such that c(r, w∗ ) = 1.
Now a isn’t an explicit parameter of our model. But for all intents and purposes it is an
exogenous primitive. To compute a all one needs is a knowledge of the production function.
As an example, recall the Cobb-Douglas case studied in Section ??, in which each training
cost corresponds to a single occupation: Cobb-Douglas form
Z X
ln y = (1 − α) ln k + α(x) ln(λ(x))dx.
0
Using the same logic as in that case, it is easy to see that a must solve the equation
Z X
(1 − α)[ln r − ln(1 − α)] + α(x)[ln(a + [1 + r]x) − ln(α(x))]dx = 0,
0
provided that condition P holds.
We are now in a position to state our central result concerning persistent inequality.
Proposition 7.12. Under [R.1], [R.2], [LP], and [P], the unique steady state is unequal if and only
if
(7.19) Ω(a, r) < a + (1 + r)X.
We shall refer to the inequality (7.19) as the widespan condition. It is made up of two parts. As
we noted above, the parameter a is entirely a feature of the production technology and the
range of the occupational structure. On the other hand, the Becker-Tomes limit wealth map
Ω is entirely a feature of preferences. The widespan condition states that Becker-Tomes limit
wealth — commencing from a — is not enough to “span” the entire range of occupations:
it (net of a and discounted by the interest rate) is smaller than the span X. Proposition 7.12
declares that in all such cases, the steady state must exhibit persistent inequality.
One should be careful enough to note that (7.19) may not prescribe a unique threshold for
the span X. Such an interpretation is suggested by the rewriting of the condition as
Ω(a, r) − a
X> ,
1+r
but observe that both a and X depend on the training cost technology. To make this a more
precise, consider economies that are identical in all respects except for their training cost
functions, which are drawn from an ordered family (all starting at 0 for some occupation).
Such an ordered family may be parameterized by the highest training cost X.
In this class, it is very easy to see that a depends negatively on X. Therefore provided
that Ω(w, r) − w is nondecreasing in w, we do generate a single-threshold restriction on span:
“there is X∗ such that widespan holds if and only if X > X∗ ”. This will be true, for instance, for
isoelastic preferences: see (7.20) and the accompanying discussion below. More generally,
though, widespan must hold for all training costs large enough (even though (7.19) may not
imply a single threshold for X).
Uniqueness plus widespan tells us that there is just one steady state, but it must treat
individuals unequally. Conversely, if widespan fails, there is convergence to a common level
of wealth for all families. The proposition asserts, therefore, that whether the disequalization
or the equalization view of the market is relevant depends on whether or not the widespan
condition is satisfied.
Observe that under widespan, the market must act to separate identical or near-identical
families, as described informally in Section ??. Long-run equality is not an option. However,
this paper does not contain an explicit account of dynamic equilibria from non-steady-state
initial conditions.
The discussion following the statement of Proposition 7.11 continues to be relevant here.
There is no history-dependence “in the large”, as the steady state is unique. But just where
an individual family will end up in that distribution profoundly depends on the distant
history of that family.
The strength of the characterization result in Proposition 7.9 is such that it renders the proof
of this proposition almost trivial. See Appendix.
7.4.8 Applications and Implications. To illustrate the implications of the span condition
in Proposition 7.12, it is useful to invoke the example of an iso-elastic utility function.
Recall, in particular, equation (7.13) and the discussion following it. If δ ≤ 1+r
1
, (7.19) is
always satisfied and an equal steady state never exists. There are effectively no financial
bequests in the limit, so the model reduces to the disequalization model in which financial
bequests are not allowed. If, on the other hand, δ ≥ 1 + 1r then Ω(a, r) = ∞ and (7.19) fails.
Financial bequests overwhelm any inequality arising from the need to provide occupational
choice incentives, and an unequal steady state cannot exist.
In the intermediate case in which δ is neither too large nor too small, (7.13) tells us that
ρ
Ω(a, r) = a,
1 − r(ρ − 1)
where ρ ≡ [δ(1 + r)]1/σ , so that the widespan condition reduces to

ρ
a < a + (1 + r)X.
1 − r(ρ − 1)
Rearranging, we obtain the following version of widespan:

ρ−1
(7.20) X>a
1 − r(ρ − 1)
We now describe effects of varying parameters of the model, which are relevant to explaining
cross-country differences, or effects of technological change.
(a) Differences in TFP Levels. Suppose we compare two countries which differ only in their
levels of total factor productivity (TFP). Then for any common value of X, the poorer country
has a lower value of a, implying that it is more prone to disequalization. Intuitively, the
lower level of wages reduces the intensity of the parental bequest motive: they are less
willing to undertake the educational investments for high-end occupations. The resulting
shortage of people in high-end occupations causes a rise in the skill premium. This elicits
the requires supply into high-end occupations, but makes wealth inequality more acute.
Technologically poorer countries are therefore more prone to disequalization.
Of course, this argument is based partly on the assumption that the range of training costs X
is unaffected by wages. However, it is easy to incorporate this extension under the plausible
assumption that both human and physical inputs enter into production. Then X lower in the
unproductive country, but not by the same factor as a. This argument is obviously reinforced
if poorer countries also possess a less productive educational technology.
(b) Differences in TFP Growth Rates. While TFP-related differences in poverty are positively
associated with disequalization, higher growth may be positively related to it as well. For
instance, if growth (from Hicks-neutral technical progress) causes all wages and costs to
grow at a uniform rate, then — all other things being equal — the level of desired bequests
will be dulled, raising the likelihood of disequalization.22
To the extent that poorer countries grow faster owing to a “catch-up” phenomenon in
technology, the widespan condition is therefore more likely to hold on two counts: higher
poverty and higher growth. Of course, the net result is ambiguous if subsequent growth
isn’t positively correlated with initial poverty.
(c) Changes in Interest Rates. A change in the rate of return to capital has subtle effects. When
r rises, ρ also goes up. Both these effects work against the widespan condition, by raising the
rate of return to financial bequests. So a first cut at this issue would suggest that an increase
in the global rate of return to physical capital tends to be equality-enhancing. However,
there is the possibility that a may be lowered by the increase in r. This effect runs in the
opposite direction, and a full analysis is yet to be conducted.
(d) Reliance on Physical Capital Now let us compare economies with differing degrees of
mechanization, i.e., reliance on physical capital vis-a-vis human capital in production. One
22We omit a formal demonstration of this assertion, which proceeds by deriving an equivalent of the widespan
condition (7.20) in the presence of neutral technical progress.
simple way to do this is to suppose that final output is produced via a nested function
y = Akα m1−α ,
where m is a composite of the occupational inputs: e.g., an intermediate good “produced”
by workers. Then greater mechanization corresponds to a rise in α. Setting the marginal
product of capital to the interest rate r, we obtain
1
k A
1−α
= ,
m r
so that the indirect “reduced-form” production function is linear in m:
y = Bm,
where
α
A
1−α
B=A .
r
Notice that B essentially prices the composite in terms of the final output. If B goes down
for some reason, then w(0) will decline. So a reduction in B, other things being equal, will
contribute to a greater likelihood of disequalization. Whether B goes up or down with α
depends on the ratio of A (TFP) to r.23 In relatively “unproductive” economies in which A
is small, an increase in physical capital intensity lowers B, making inequality more likely.
The opposite is the case in “productive” economies in which A is large. We thus obtain an
interesting answer to a classic question in the theory of distribution: the impact of greater
mechanization in production on long-run inequality.
(e) Wider Product Variety. Wider occupational spans may be the outcome of introduction
of new goods and services, owing to technological change. The production of new goods
and services such as information and communication technology creates an entirely new
set of occupations. Such occupations are likely to require high levels of education and
training, which may be thought of as an increase in the span of occupations and associated
training costs. Unlike the parameterization used in Proposition 7.12, such changes involve
an increase both in X and in the productivity of the technology. In terms of (7.20), both X
and a tend to rise and the net effect depends on the ratio of these two variables.
We have not yet analyzed this application in detail, but it is clear that the effect of wider
product variety on inequality is a very important question.
(f) When (7.19) holds, the rising rate of return to occupations (as captured by Observation
7.5) is an important prediction of the theory. We have discussed this already, but it bears
repetition that this derived property of the model contradicts the assumed properties of human
capital in Becker and Tomes (1986) and others. Empirical research will take us closer to
settling this issue.
23The intuitive reason why this is the appropriate comparison is that A becomes the productivity of capital in
the limiting case when α = 1.
7.4.9 A Summary. We’ve addressed two central questions in the theory of income distri-
bution:
[I] Do competitive markets equalize or disequalize wealth allocations?
[II] Does history matter? Are the same economic fundamentals consistent with multiple
steady states?
We study a model of intergenerational bequests which allows for both financial bequests
as well the choice of a rich variety of occupations. A fundamental postulate of the model
is that occupational inputs are imperfect substitutes, so that factor prices are endogenously
determined. At the level of an individual household, occupational investments may be
fine-tuned to an arbitrary degree, provided that there is rich variation in occupations and
training costs. But the returns to these occupations are endogenous, and the equilibrium
of the market will determine whether households face a convex or nonconvex investment
technology.
We provide a complete characterization of steady states. We show that if there is inequality
in steady state (so that question I has been answered affirmatively), then the steady state
wage function must be linear over a section and convex over others, so that each family
must face an investment nonconvexity. This finding is at sharp odds with existing literature
that simply assumes the opposite: that the rate of retrun to human capital must exhibit
diminishing returns.
Our characterization permits us to move on to the two central questions. We prove that with
a rich set of occupations, the steady state must be unique. There may (and generally will)
be path-dependence at the level of dynastic choices, but the overall distribution of outcomes
must be independent of history. This provides an interesting negative answer to question
II, and in this way challenges a literature that is predicated on steady state multiplicity.
The second principal goal is to address question I: we extend and unify three seemingly
different views of market-driven inequality: equalization, embodied in traditional theories of
income distribution, disequalization, central to the recent endogenous inequality literature,
and neutrality, that either can happen depending on historical conditions. We show that a
fundamental condition, which we call the widespan condition, allows us to predict whether
the steady state must involve inequality or not.
The widespan condition draws attention to an aspect of technology that has received little
attention in the literature: the range or “span” of occupational structure. Whether equality
or inequality results from market mechanisms depends on this key parameter representing
the extent of occupational diversity, relative to the strength of bequests.
The theory has several potential applications and implications.
We end with a final comment on question II. Our uniqueness result stems from the assumed
richness in occupational investments. Whether or not such richness exists therefore deserves
to be studied empirically. If the assumption is valid, theories of macroeconomic history
dependence will have to be based either on interest rate multiplicities rooted in the lack
of capital market integration, or political economy channels, rather than market-based

occupational choice mechanisms.
7.4.10 Proofs. Proof of Proposition 7.7. Define κ̄(λ) ≡ u(w̄(λ)) − u(w̄(λ) − x); this is the utility
cost of acquiring education when the parent is high. Define a similar utility cost for the low
parent: κ(λ) ≡ u(w(λ)) − u(w(λ) − x). Let b(λ) ≡ u(w̄ − x) − u(w) be the one-period gain to
¯
¯ (assuming
being high ¯ high parent also invests in her child
that the ¯ and the low parent does
not), and define B(λ) ≡ (1 − δ) b(λ).
−1
Finally, for any sequence {λs } and for any date t, define
X∞
Bt ≡ δs−t b(λs ).
s=t
This is the lifetime gain between a currently high and a currently low dynasty (starting from
any date t), assuming that dynasties never switch their skill status.
The reason why Bt acquires salience is given by the following simple observation, which
states that at every date, the equilibrium lifetime utility of the high (and low) must be equal
to the utility they would have received were their descendants never to switch status. [To be
sure, along the equilibrium path, switching of status will generally occur nevertheless.]
Lemma 7.1. If {w̄t , wt , λt }∞
t=0
is a competitive equilibrium, then for each date t,
¯
X∞
(7.21) V̄ t = δs−t u(w̄s − x)
s=t
and
∞
X
(7.22) Vt = δs−t u(ws ),
¯ s=t
¯
so that in particular,
(7.23) V̄ t − V t = Bt for all t.
¯
Proof. It suffices to show that for each t ≥ 0,

V̄ t = u(w̄t − x) + δV̄ t+1
and
V t = u(wt ) + δV t+1 .
¯ ¯ ¯
To prove this, apply Observations 7.1 and 7.2. By Observation 7.1 and our restriction on
λ0 , λt ∈ (0, λ̃) for all t ≥ 0. Now using Observation 7.2, we may conclude that at all dates,
some of the high people stay high, while some of the low people stay low. This is enough to
establish the result.
Thus along any competitive equilibrium, no dynasty will strictly prefer to switch skills,
though it may well be the case that it strictly prefers to stay where it is. Lemma 7.1 yields,
in turn

t=0
is a competitive equilibrium, then for every t, (w̄t , wt ) are the wages
¯
associated with λt , and ¯
(7.24) κ̄(λt ) ≤ δBt+1 ≤ κ(λt ),

¯
with
(7.25) λt+1 > λt only if δBt+1 = κ(λt )
¯
and
(7.26) λt+1 < λt only if δBt+1 = κ̄(λt )
Proof. Let {w̄t , wt , λt } be a competitive equilibrium. Then by definition, (w̄t , wt ) must be the
wages associated ¯ with λ for every t. Using (7.21) and utility maximization, we ¯ see that
t
∞
X
V̄ t = δs−t u(w̄s − x) ≥ u(w̄t ) + δV t+1 ,
s=t
¯
so that (using (7.21) again)
u(w̄t ) − u(w̄t − x) ≤ δ[V̄ t+1 − V t+1 ]
¯
with equality holding whenever a switch from “high” to “low” does occur along the
equilibrium path. Invoking (7.23) of Lemma 7.1, we get half of (7.24) as well as (7.26).
The same argument applied to a currently low dynasty gets us the other half of (7.24) and
(7.25).
The next step is central:

t=0
is a competitive equilibrium, then for every t,
¯
1 1
(7.27) max{B(λt ), κ̄(λt )} ≥ Bt+1 ≥ min{B(λt ), κ(λt )}.
δ δ¯
Proof. It suffices to prove the result for t = 0. I first show that

1
(7.28) max{B(λ0 ), κ̄(λ0 )} ≥ B1
δ
Suppose this assertion is false. Then I claim that there exists a first date T ≥ 0 such that
λ0 = · · · = λT and
1
(7.29) BT+1 ≥ κ(λT ).
δ¯
Of course, if B1 ≥ (1/δ)κ(λ0 ), the claim is automatically true; otherwise B1 < (1/δ)κ(λ0 ). We
also have (by virtue of ¯the presumption that (7.28) does not hold) that B1 > (1/δ)¯ κ̄(λ0 ), so
that
1 1
κ(λ0 ) < B1 < κ̄(λ0 )
δ¯ δ
which — by Lemma 7.2 — implies that λ0 = λ1 . Moreover, B1 = b(λ1 ) + δB2 = b(λ0 ) + δB2 ,
while B(λ0 ) = b(λ0 ) + δB(λ0 ), so that
1
B2 − B(λ0 ) = [B1 − B(λ0 )],
δ
and simple manipulation of this equality shows that
1−δ
(7.30) B2 = B1 + ,
δ
where ≡ B1 − B(λ0 ) > 0 (again, by the failure of (7.28)).
Following the same reasoning leading up to (7.30), as long as Bt+1 < (1/δ)κ(λt ) — and as
¯
long as this is also true of all dates before t — we have λ0 = · · · = λt = λt+1 , and
1−δ
(7.31) Bt+1 = Bt + ,
δ
It follows that there must be a first date T such that λ0 = · · · = λT , and (7.29) holds, as
claimed.
On the other hand, (7.29) cannot hold with strict inequality, as this would surely violate
(7.24) of Lemma 7.2, so it must be that
1
(7.32) BT+1 = κ(λT ).
δ¯
In turn, this means that λT+1 ≥ λT (see (7.26)). Moreover, δBT+2 = BT+1 − b(λT+1 ), so that
δBT+2 − κ(λT+1 ) = BT+1 − b(λT+1 ) − κ(λT+1 )
¯ ¯
≥ BT+1 − b(λT ) − κ(λT )
¯
= BT+1 − b(λT ) − δBT+1
(7.33) = BT+1 − BT > 0,
which contradicts (7.24) at date T + 1. This establishes (7.28).
Now we prove that
1
(7.34) B1 ≥ min{B(λ0 ), κ(λ0 )}
δ¯
The argument runs closely parallel to the previous one. Suppose (7.34) is false. Then I claim
that there exists a first date T ≥ 0 such that λ0 = · · · = λT and
1
(7.35) BT+1 ≤ κ̄(λT ).
δ
The steps are very similar to those used to establish (7.29), and are omitted.
Continuing the parallel argument, (7.35) cannot hold with strict inequality, so we have
1
(7.36) BT+1 = κ̄(λT ).
δ
In turn, this means that λT+1 ≤ λT . These two observations establish, however, that
δBT+2 − κ̄(λT+1 ) < 0,
(following steps parallel to those establishing (7.33)), which contradicts (7.24) and completes
the proof of the lemma.
The following observation is a simple consequence of Lemma 7.3:

Lemma 7.4. Along any equilibrium, if B(λt ) ≤ (1/δ)κ̄(λt ), then Bt+1 = (1/δ)κ̄(λt ). Similarly, if
B(λt ) ≥ (1/δ)κ(λt ), then Bt+1 = (1/δ)κ(λt ).
¯ ¯
Proof. We prove the first part; the second part uses a completely analogous argument. If
B(λt ) ≤ (1/δ)κ̄(λt ), then max{B(λt ), (1/δ)κ̄(λt )} = (1/δ)κ̄(λt ), so that by (7.27), (1/δ)κ̄(λt ) ≥
Bt+1 . On the other hand, Lemma 7.2 tells us that (1/δ)κ̄(λt ) ≤ Bt+1 , and the proof is complete.
With these steps in hand, we may complete the proof of the theorem. There are three
possibilities to consider (each a restriction on the initial value λ0 ):
I. (1/δ)κ̄(λ0 ) < B(λ0 ) < (1/δ)κ(λ0 ). Then by (7.27), B(λ0 ) = B1 . In particular,

¯
(1/δ)κ̄(λ0 ) < B1 < (1/δ)κ(λ0 )
¯
so that by Lemma 7.2, λ0 = λ1 . Continuing the argument recursively, we see that λt = λ0 for
all t.
II. B(λ0 ) ≤ (1/δ)κ̄(λ0 ). Then by Lemma 7.4, B1 = (1/δ)κ̄(λ0 ) and so by Lemma 7.2, λ1 ≤ λ0 .
Suppose, in fact, that strict inequality holds. Then (1/δ)κ̄(λ1 ) < (1/δ)κ̄(λ0 ) and (1/δ)κ(λ1 ) >
(1/δ)κ(λ0 ), so that ¯
¯
(7.37) (1/δ)κ(λ1 ) > B1 > (1/δ)κ̄(λ1 ).
¯
Now I claim that in fact,
(7.38) (1/δ)κ(λ1 ) > B(λ1 ) > (1/δ)κ̄(λ1 ).
¯
Suppose not. First suppose that B(λ1 ) ≤ (1/δ)κ̄(λ1 ). Then by Lemma 7.4, B2 ≤ (1/δ)κ̄(λ1 ), so
that B1 = (1 − δ)B(λ1 ) + δB2 ≤ (1/δ)κ̄(λ1 ), which contradicts (7.37). In exactly the same way,
one can rule out the possibility that B(λ1 ) ≥ (1/δ)κ(λ1 ), so (7.38) is established.
¯
Now we are in Case I, and λ must remain constant thereafter. So in Case II, we move to a
steady state in at most one step.
III. B(λ0 ) ≥ (1/δ)κ(λ0 ). Then by Lemma 7.4, B1 = (1/δ)κ(λ0 ) and so by Lemma 7.2, λ1 ≥ λ0 .
¯ ¯
To bring out the contrast with Case II, I claim that B(λ1 ) ≥ (1/δ)κ(λ1 ), with strict inequality
if the corresponding inequality at date 0 also holds strictly. ¯
Suppose on the contrary that B(λ1 ) < (1/δ)κ(λ1 ). Then

¯
(7.39) B1 = (1/δ)κ(λ0 ) ≥ (1/δ)κ(λ1 ) > B(λ1 ),
¯ ¯
while it is also true that
(7.40) B1 = (1/δ)κ(λ0 ) ≥ (1/δ)κ(λ1 ) ≥ B2 .
¯ ¯
But (7.39) and (7.40) together contradict the fact that B1 = (1 − δ)B(λ1 ) + δB2 .
If strict inequality holds at date 0 — B(λ0 ) > (1/δ)κ(λ0 ) — then we can arrive at a contradiction
simply under the weaker condition B(λ1 ) ≤ (1/δ) ¯ κ(λ ). For then λ , λ and therefore
1 1 0
(because λ1 ≥ λ0 ) it must be that λ1 > λ0 . Therefore ¯ the weak inequality in (7.39) holds
strictly, and we obtain the same contradiction.
So the claim is established and we can apply Case III repeatedly to argue that λt+1 ≥ λt for
all t in this case.
Moreover, if B(λ0 ) > (1/δ)κ(λ0 ), then B(λt ) > (1/δ)κ(λt ) for all t subsequently, so convergence
¯ in contrast to the “one-step”
cannot ever occur in finite time, ¯ property of Case II.
Finally, observe that in the case B(λ0 ) > (1/δ)κ(λ0 ), in which λt+1 > λt for all t, there is no
t and no λ ∈ [λt , λt+1 ] which is a steady state.¯ For suppose there were; then in particular,
κ(λ) ≥ B(λ), so that
¯
(7.41) κ(λt ) ≥ κ(λ) ≥ B(λ) ≥ B(λt+1 ) > Bt+1 ,
¯ ¯
where the very last inequality follows from the fact that λs < λs+1 for all s. But (8.42)
contradicts the equilibrium condition (7.24). This proves that in Case III, convergence
occurs to the smallest steady state to the right of λ0 .
Proof of Observation 7.3. Let Wt denote the wealth of a typical member of generation t. We
claim that if Wt+1 < Wt , Ws+1 ≤ Ws for all s > t. Suppose not; then there exist dates τ
and s with s > τ such that (a)Wτ+1 < Wτ , (b) Ws+1 > Ws , and (c) Wm = Wτ+1 = Ws for all
intermediate dates τ+1 ≤ m ≤ s (this last requirement is, of course, vacuous in case s = τ+1).
But then a strictly higher wealth (Wτ compared to Ws ) generates a strictly lower descendant
wealth (Wτ+1 compared to Ws+1 ), which contradicts a familiar “single-crossing” argument
based on the strict concavity of the utility function. Therefore dynastic wealth converges in
this case.
Next suppose that Wt+1 > Wt . Then a reversal of the same logic implies that family wealth
is nondecreasing across generations.
Lemma 7.5. In the Becker-Tomes benchmark, W̃ − W has exactly the same sign (including equality)
as
U0 ( rW+w
1+r )
(7.42) (1 + r) − .
V 0 (W)
Proof. Simply recall the first-order conditions (9.42) and the concavity of U and V.
Proof of Observation 7.4. We first show that under (H),

rW + w
(7.43) U0 ( )/V 0 (W) is increasing in W,
1+r
for all w > 0 and r > −1. It is sufficient to show that
r rW + w rW + w 00
V 0 (W) U00 ( ) − U0 ( )V (W) > 0
1+r 1+r 1+r
which is equivalent to
U00 ( rW+w
1+r )

r

V 00 (W)
− < − .
U0 ( rW+w
1+r )
1+r V 0 (W)
Given (H), this reduces to the condition that
r α + β[ rW+w
1+r ]
< ,
1+r α + βW
which is always true unless w = 0 and α = 0. This immediately verifies [LP], barring these
exceptional cases.
If w = 0 and α = 0, then U0 ( rW+w 0
1+r )/V (W) is unchanging in W, so that [LP] is verified for all
but one value of r.
To prepare for the proofs of the remaining propositions, we record several lemmas, and we
presume (often implicitly) that [R], [E] and [LP] apply where needed.
Lemma 7.6. In the Becker-Tomes benchmark, for every w > 0 and r > −1:
(a) If w ≤ W < Ω(w, r), then W < W̃(W; w, r) ≤ Ω(w, r). If W > Ω(w, r), then Ω(w, r) ≤
W̃(W; w, r) < W.
(b) Ω(w, r) is nondecreasing in w.
(c) Ω(w, r) is continuous at every (w, r) at which it is finite.
Proof. Part (a) follows simply from the fact that W̃(W; w, r) is nondecreasing in W. If, for
instance, W̃(W; w, r) ≤ W < Ω(w, r), or if W̃(W; w, r) > Ω(w, r) > W, limit wealth starting from
W can never equal Ω(w, r). This proves the first assertion in part (a); the second assertion
follows in similar fashion.
Part (b). We claim first that W̃(W; w, r) is nondecreasing in w. To show this, it suffices to
assume that W̃ > w, but then it must be the case that
!
W̃ − w
(7.44) 0
U W− ≤ (1 + r)V 0 (W̃).
1+r
0
Now suppose that w goes up to w0 , but W̃ falls to W̃ . The new situation must therefore
have strictly positive consumption for the parent, so that the new first-order condition has
the opposite weak inequality:
0
W̃ − w0
!
0
(7.45) 0
U W− ≥ (1 + r)V 0 (W̃ ).
1+r
But (7.44) and (7.45) together contradict the assumption that U and V are strictly concave.
With this claim in place, the result follows immediately from [LP].
Part (c). Suppose that Ω(w, r) is finite. By part (a), W̃(W; w, r) (viewed as a function of W)
must “strictly intersect the 450 line” at the value Ω(w, r). The continuity of W̃(W; w, r) now
assures us that Ω(w, r) must be locally continuous.
Lemma 7.7. For any steady state wage function:
(a) If h is inhabited, x(h) = x(h0 ) implies w(h) ≥ w(h0 ).
(b) If h is inhabited, then x(h) > x(h0 ) implies w(h) − w(h0 ) ≥ (1 + r)[x(h) − x(h0 )].
Proof. If (a) is false and x(h) = x(h0 ) while w(h) < w(h0 ) then any parent selecting occupation
h for her child would do better to select occupation h0 instead. The same is true if (b) were
false: the parent could switch to occupation h0 for the child, combined with a higher financial
bequest so as to leave the child’s wealth unaffected, while increasing her own consumption.
As in the main text, it will often be convenient to write the wage as a function of training
cost: w(x). Part (a) of Lemma 7.7 informs us that we can certainly do this right away for
inhabited training costs.
Lemma 7.8. (a) The set of all inhabited training costs — call it T — is a subset of [0, X] of full
measure, and w is continuous on T.
(b) There is a unique continuous extension of w on T to all of [0, X], and it forms an equivalent
representation.
Proof. Part (a). Let T be the set of all inhabited training costs. Since a steady state must
have positive output by definition, it follows from [R] and [E] that T must be of full measure.
Moreover, w (viewed as a function of x) must be continuous on T. For if not, we can select
training costs x and x0 in T that are arbitrarily close, but such that their wage difference is
bounded away from zero. In that case no parent would select the (almost identical) training
cost with a lower wage.
Part (b). Because w is nondecreasing on T and T is of full measure, there is a unique
continuous extension of w(x) to all of [0, X]. Use this extension to define ŵ(h) for all
occupations, inhabited or not. That is, ŵ(h) = w(h) if h is inhabited, and equals the continuous
extension otherwise. We first claim that ŵ(h) ≥ w(h) for all h that are uninhabited. For if not,
we have a contradiction in a manner similar to part (a): we can find an inhabited occupation
h0 arbitrarily close to h but with wages bounded below that of w(h), which means that all
occupiers of h0 would prefer h, a contradiction.24
By this claim, if we replace w by ŵ, no firm will wish to change its desired input mix (unused
inputs have not become any cheaper). To complete the proof of equivalence, observe that no
family occupying h0 finds it strictly profitable to switch to an uninhabited occupation h once
its wage has been replaced by ŵ(h). For if this were true, then by the definition of continuous
extension we can find a third inhabited occupation h00 with wage and training cost arbitrarily
close to that of h, such that the family must therefore also find it profitable to switch from h0
to h00 . But this is a contradiction, since that option is already available in the going steady
state.
24This argument, as well as its counterpart for (a), can be made entirely precise by showing that the preference
for h over h0 can be made uniform over all families, irrespective of their wealth.
In what follows we focus exclusively on this equivalent representation and call it w instead
of ŵ. Define w = w(0).
Lemma 7.9. Every family attains a wealth of at least Ω(w, r).
Proof. Let W be the steady state wealth of some family. Suppose that it incurs training
cost x and leaves bequest b; then by stationarity, W = (1 + r)b + w(x). Now, given the choice
of x, b must maximize
(7.46) U(W − x − b) + V(w(x) + (1 + r)b)
subject to b ≥ 0. Defining B ≡ b + x, this can equivalently be written as: B is chosen to
maximize
(7.47) U(W − B) + V(w0 + (1 + r)B)
subject to B ≥ w(x), where w0 ≡ w(x) − (1 + r)x. It is obvious that the solution to this problem
involves a total bequest B at least as large as the value that would obtain if the constraint
B ≥ w(x) were replaced by B ≥ 0, i.e., if we were in a Becker-Tomes benchmark world with w0
and r. Remembering that W is also next period’s wealth in the problem (7.47), we conclude
that
(7.48) W ≥ W̃(W; w0 , r).
We now claim that

(7.49) W ≥ Ω(w0 , r).
If this is false, then W < Ω(w0 , r). Combining this with (7.48), we may conclude that
W̃(W; w0 , r) ≤ W < Ω(w0 , r), but this contradicts part (a) of Lemma 7.6.
Now recall the definition of w0 and invoke part (b) of Lemma 7.7 to conclude that w0 ≥ w.
Use this fact and part (b) of Lemma 7.6 to observe that Ω(w0 , r) ≥ Ω(w, r). Combine this last
inequality with (7.49) to obtain the desired result.
Lemma 7.10. Let W be any lower bound on stationary wealth across all families.
(a) If for any occupation h, we have w + (1 + r)x(h) ≤ W, then w(h) = w + (1 + r)x(h).
(b) The stationary wealth of any family selecting an occupation h with w + (1 + r)x(h) ≤ W must be
Ω(w, r).
Proof. Part (a). Suppose, on the contrary, that w + (1 + r)x(h) ≤ W, but w(h) , w + (1 + r)x(h).
By Lemma 7.7, part (b), and the fact that w is a continuous equivalent representation, it must
be that
w(h) > w + (1 + r)x(h).
so that for some x < x(h) but close to it,
0
(7.50) w(x0 ) > w + (1 + r)x0 .
Now, we know that there is a sequence of inhabited occupations {hk } such that x(hk ) ↓ 0. By
assumption, W is a lower bound on stationary wealth across all families, and w + (1 + r)x0 <
w + (1 + r)x(h) ≤ W. Therefore (for k large enough) such families are almost exclusively
leaving financial bequests of magnitude at least x0 to their children at rate of return r. They
would be better off, instead, selecting an occupation with cost x0 for their children (which
yields a return of strictly more than r, by (7.50), and supplementing the remainder with
financial bequests, a contradiction.
For each family selecting occupation h with w + (1 + r)x(h) ≤ W, we have w(h) = w + (1 +
r)x(h). Therefore the realized rate of return to all the choices of such a family, financial and
educational, is exactly r. Define x̂ by
(7.51) w + (1 + r)x̂ = W.
Once again, using Lemma 7.7, part (b), and the fact that w is a continuous equivalent
representation, we also know that wages for training costs beyond x̂ yield no less a return
than r for all educational investments beyond x̂. Yet these families choose (by part (a)) not to
utilize such regions of educational investment. They must therefore be behaving in exactly
the same way as in a Becker-Tomes benchmark world with parameters (w, r). We must
conclude that their stationary wealth equals Ω(w, r).
Lemma 7.11. In an equal steady state, the common wealth of all families must be Ω(w, r), and
Ω(w, r) ≥ w + X(1 + r).
Proof. Let W denote the (common) wealth of each family in an equal steady state. We
claim that
(7.52) W ≥ w + (1 + r)X.
For if not, we know from Lemma 7.9, part (b) (and the fact that w is a continuous equivalent
representation) that there are inhabited occupations with training costs x arbitrarily close to
X, and that
w(x) ≥ w + (1 + r)x.
Consequently, if (7.52) were to be false, we would find inhabited occupations with w(h) > W.
Because the wealth of families in such occupations is at least w(h), this is a contradiction to
equality.
It remains to prove that W = Ω(w, r). Because (7.52) is true, and because W is (trivially)
a lower bound on stationary wealth, Lemma 7.10, part (b) applies to all families, and
W = Ω(w, r).
Lemma 7.12. A steady state is equal if and only if the continuous equivalent representation of the
wage function is linear:
(7.53) w(h) = w + (1 + r)x(h)
for all occupations h.
Proof. For necessity, combine Lemma 7.10, part (a), and Lemma 7.11. For sufficiency, note
that if (7.53) holds, then we are in a Becker-Tomes benchmark world and all limit wealths
must be the same.
The remaining steps concern unequal steady states. Define θ by

(7.54) w + (1 + r)θ = Ω(w, r).
Lemma 7.13. In an unequal steady state, it must be that θ < X.
Proof. By Lemma 7.9, Ω(w, r) is a lower bound on stationary wealth for all families. If,
contrary to our assertion, θ ≥ X, then part (a) of Lemma 7.10 applies for every training cost,
when W is replaced by Ω(w, r). Therefore w satisfies (7.53), and the sufficiency direction of
Lemma 7.12 implies that the steady state must be equal, a contradiction.
We are interested in the shape of w in the region U ≡ [θ, X].

Lemma 7.14. Let I be some subinterval of U such that no financial bequests are made by any family
that occupies some occupation with training cost in U. Then w satisfies (7.17) on I.
Proof. Fix any x ∈ I, with x < sup I. For > 0 but small enough, x + ∈ I as well.
Assume provisionally that both x and x + are inhabited. Then family wealth at x (resp.
x + ) is merely w(x) (resp. w(x + )). Using the two optimality conditions, one for families
with wealth w(x) and the other for families with wealth w(x + ), we see that
(7.55) U(w(x)−x)−U(w(x)−(x+)) ≥ V(w(x+))−V(w(x)) ≥ U(w(x+)−x)−U(w(x+)−(x+)).
Now, using the fact that w is a continuous equivalent representation, and invoking [R] and
[E], we can see that (7.55) actually applies to all x and x + in I, not just those that are
inhabited.25
Dividing these terms throughout by , applying the concavity of the utility function to the
two side terms, and the mean value theorem to the central term, we see that
w(x + ) − w(x)
U0 (w(x) − (x + )) ≥ V 0 (γ()) ≥ U0 (w(x + ) − x),

where γ() lies between w(x) and w(x + ). Now send to zero and use the continuous
differentiability of U and V to conclude that the right-hand derivative of w with respect to x
— call it w+ (x) — exists, and
w+ (x) = U0 (w(x) − x)/V 0 (w(x)).
By exactly the same argument applied to x (greater than inf I) and x − , we may conclude
the same of the left-hand derivative, which verifies (7.17).
Lemma 7.15. Any two-phase wage function has w0 (x) > 1 + r for almost all x > θ.
Proof. The continuous differentiability of U and V, and the continuity of w together

imply that w is continuously differentiable in its second phase, where it follows (7.17). Note
moreover that w0 (θ) = 1 + r. Therefore, if the assertion is false, there is an interval [x1 , x2 ],
with x1 ≥ θ, on which w0 (x) ≤ 1 + r, while w0 (x1 ) = 1 + r and w(x1 ) − (1 + r)x1 ≥ 0.26 Applying
25Given [R] and [E], we can approach both x and x + by a sequence of inhabited training cost pairs in I, and
for each such pair the inequality (7.55) holds.
26One way to assure the existence of an interval with all these properties is to take x to be the minimum of the
1
values among those greater than θ for which w0 (x) ≤ 1 + r.
(7.17) at x1 and using w0 (x1 ) = 1 + r, we see that

(7.56) U0 (w(x1 ) − x1 )) = (1 + r)V 0 (w(x1 )).
Define ŵ = w(x1 ) − (1 + r)x1 , which is nonnegative by construction. Then (7.56) reduces to
0 rw(x1 ) + ŵ
!
(7.57) U = (1 + r)V 0 (w(x1 )),
1+r
By the same logic as (7.56), except that at x2 we have w0 (x2 ) ≤ 1 + r, we see that
(7.58) U0 (w(x2 ) − x2 )) ≤ (1 + r)V 0 (w(x2 )).
Now observe that w(x1 ) + (1 + r)(x2 − x1 ) ≥ w(x2 ) (because w0 (x) ≤ 1 + r on [x1 , x2 ]), or
w(x2 )−ŵ
equivalently, using the definition of ŵ, x2 ≥ 1+r . Consequently,
rw(x2 ) + ŵ
w(x2 ) − x2 ≤ ,
1+r
and using this in (7.58) along with the concavity of U, we must conclude that
0 rw(x2 ) + ŵ
!
(7.59) U ≤ (1 + r)V 0 (w(x2 )).
1+r
Suppose that w(x1 ) > 0. Then (7.57) means that w(x1 ) is a positive limit wealth in the Becker-
Tomes benchmark with (ŵ, r), while Lemma 7.5 tells us that W̃(w(x2 ); ŵ, r) ≥ w(x2 ). Because
w(x2 ) > w(x1 ), this means that w(x1 ) cannot be a limit wealth starting from initial wealth
w(x2 ). This contradicts [LP].
On the other hand, if w(x1 ) = 0, then w = 0 and θ = 0 as well, which means that Ω(0, r) = 0:
the limit of Becker-Tomes wealth from all positive initial wealths is zero when w = 0.27 But
this fact is contradicted by Lemma 7.5 applied to (7.59), because ŵ is also 0 in this case.
If a wage function satisfies w(x) − w(x0 ) = (1 + r)(x − x0 ) for all x and x0 in some interval, say
that it is r-linear over that interval. We know, for instance, that any two-phase wage function
with a nondegenerate first phase indeed r-linear over [0, θ(w)].
Lemma 7.16. Suppose that a family in steady state, inhabiting training cost x at some date, also
makes a financial bequest at that date. That is, it possesses (and bequeaths) total wealth W, where
W > w(x). Then w is r-linear over all x0 ≥ x with w(x) + (1 + r)(x0 − x) ≤ W:
(7.60) w(x0 ) = w(x) + (1 + r)(x0 − x).
Proof. The proof is very similar to that of Lemma 7.10. Pick any x0 with w(x) + (1 + r)(x0 −
x) ≤ W. Our family is making a financial bequest of at least x0 − x. If (7.60) were to fail, then
by Lemma 7.7, part (b), and the fact that w is a continuous equivalent representation, we
must have
w(x0 ) > w(x) + (1 + r)(x0 − x),
which means that our family would certainly be better off choosing x0 instead of x and
supplementing the remainder (if any) with financial bequests, a contradiction.
27Remember: this is a different statement from the one that asserts that 0 is a Becker-Tomes limit wealth when
w = 0, which is always trivially true.
w*
w (x) = w* (x')
x′ x x
Figure 7.7. Proof of Lemma 7.18
The next lemma summarizes what we know so far about an unequal steady state.
Lemma 7.17. The continuous equivalent-representation of any unequal steady state wage function
is r-linear up to θ, followed by combinations of intervals over which either the differential equation
(7.17) is obeyed, or r-linearity holds.
Proof. Combine Lemmas 7.14 and 7.16.
However, we now establish a stronger property:

Lemma 7.18. In an unequal steady state, the continuous, equivalent-representation wage function
must be two-phase.
Proof. Let w be a (continuous) steady state wage function, starting from w. Denote by
w∗ the two-phase wage function starting from the same point. We know already that the
two functions coincide at least up to θ. Suppose, contrary to the assertion, that w∗ (x) , w(x)
for some x ∈ (θ, X]. Then there is some first r-linear segment “after” θ at which w departs
from w∗ .
By Lemma 7.15, w must lie below w∗ in this segment. Use Figure 7.7 as a guide in what
follows.
Pick some inhabited x in the interior of the r-linear segment; then pick x0 ∈ (θ, x) such that
(7.61) w∗ (x0 ) = w(x) > 0.
Pick any family that inhabits x at any date, and has stationary wealth W. The first-order
conditions for utility maximization tell us that
rW + w
U0 ( ) = (1 + r)V 0 (W),
1+r
so that W > 0 is a limit wealth in the Becker-Tomes benchmark with baseline wage w0 =
w(x) − (1 + r)x and rate of return r. By [LP] applied to this benchmark, a family with starting
wealth W 0 = w(x) > 0 in this benchmark world must converge to the very same limit wealth.
Because W 0 = w(x) ≤ W, convergence to W must require that W̃(W 0 ; w0 , r) ≥ W 0 . Invoking
Lemma 7.5 and writing w(x) in place of W 0 , we must conclude that
rw(x) + w0
(1 + r)V 0 (w(x)) ≥ U0 ( ) = U0 (w(x) − x).
1+r
Now recall the definition of x0 from (7.61). Replacing w(x) by the same value w∗ (x0 ), replacing
x by the smaller value x0 , and using the strict concavity of U, we see that
(1 + r)V 0 (w∗ (x0 )) > U0 (w∗ (x0 ) − x0 ),
but this contradicts the fact that w∗ satisfies (7.17) at x0 .
Proof of Proposition 7.9. Lemma 7.8 establishes that there is a continuous equivalent
representation to the wage function in every steady state. Lemma 7.11 shows that in any
equal steady state, Ω(w, r) ≥ w + (1 + r)x, where w is the lowest wage in that steady state
wage function. Lemma 7.53 shows that the wage function must be r-linear for equal steady
states. Lemma 7.18 shows that unequal steady state wage functions must have the two-
phase property: it is r-linear up to θ, which is defined in (7.54), and follows the differential
equation (7.17) thereafter. Lemma 7.15 shows that the second phase must exhibit a rate of
return that is almost everywhere higher than r.
Proof of Observation 7.5. The differential equation (7.17) in the exponential utility case reduces
to
1
(7.62) w0 (x) = exp(αx)
δα
from which the stated result follows. Applying (7.17) to the constant elasticity case, we see
that for all x ≥ θ,
" #σ
1 w(x)
(7.63) w (x) =
0
.
δ w(x) − x
Differentiation of this equality shows us that
#σ−1
w(x) − xw0 (x)
"
w(x)
w (x) = σ
00
,
w(x) − x [w(x) − x]2
so that w00 (x) is continuous and has precisely the same sign as w(x)/x − w0 (x). Notice that
w(x)
> w0 (x)
x
at x = θ. So w0 (x) increases just to the right of θ, while — using (7.63) — w(x)/x monotonically
falls. But it must be the case throughout that w(x)/x continues to exceed w0 (x), otherwise the
very changes described in this paragraph cannot occur to begin with. Therefore w0 (x) rises
throughout, establishing strict convexity to the right of θ.
However, w0 cannot go to ∞, as another perusal of (7.63) will readily reveal. Indeed, w0

converges to a finite limit, which is computed by setting both w0 (x) and w(x)/x equal to the
same value in (7.63).
Proof of Proposition 7.10. The necessity of [P] is obvious, given the characterization in
Proposition 7.9, so we establish sufficiency.
Index each two-phase wage function w by its starting wage w, and define c∗ (w) = c(r, w).
Condition P assures us that c∗ (w1 ) ≤ 1 for some w1 . We claim that c∗ (w2 ) > 1 for some w2 .
Suppose not; then c∗ (w) ≤ 1 for all w. Send w ↑ ∞, then to maintain c∗ (w) ≤ 1 it must be
that the associated cost-minimizing λ — call it λ(w) — converges weakly to 0. Fix any k > 0.
Then for w large enough, [E] implies that
f (k, λ(w))/k < r.
For all such w, concavity of f in k tells us that the associated cost-minimizing capital input
k(w) must be bounded. But now the continuity of f (together with [E]) atells us that output
goes to zero as w → ∞, which contradicts unit cost minimization. This proves the claim.
Because c∗ is continuous,28 there exists w∗ between w1 and w2 such that c∗ (w∗ ) = 1.
We prove that the two-phase wage function w emanating from w∗ satisfies all the conditions
for a steady state wage function. To this end, we specify a steady state wealth and bequest
distribution, and occupational choice.
First, let λ∗ be the input mix associated with the supporting wage function w. Arrange the
population over occupations according to λ∗ . Let θ = θ(w∗ ).
If a family i is assigned to occupation h with x(h) ≤ θ, set that family’s wealth equal to Ω(w∗ , r),
its educational bequest equal to x(h), and its financial bequest equal to [Ω(w∗ , r) − x(h)]/(1 + r).
Otherwise, if occupational assignment h has x(h) > θ, set that family’s wealth equal to
w∗ (x(h)), its educational bequest equal to x(h), and its financial bequest equal to 0.
To complete the proof, we must show that each family chooses an optimal bequest. First pick
a family located at occupation h with x = x(h) ≥ θ. Because w∗ has a slope of at least 1 + r in x,
this family has no need to make financial bequests. Let M(x, x0 ) ≡ U(w(x) − x0 ) + V(w(x0 )) be
this family’s expected payoff from leaving an educational bequest x0 , and let N(x) ≡ M(x, x).
Then N is differentiable and it is easy to see that
(7.64) N0 (x) ≥ U0 (w(x) − x)w0 (x) for all x, with equality if x ≥ θ.
28This is a consequence of the maximum theorem and the assumption that production is continuous in the weak
topology over occupational distributions on H.
For any x0 ≥ x ≥ θ, then, using the equality in (7.64)

Z x0
(7.65) M(x, x) = M(x , x ) −
0 0
U0 (w(z) − z)w0 (z)dz
x
Z x0
(7.66) ≥ M(x0 , x0 ) − U0 (w(z) − x0 )w0 (z)dz
x
(7.67) = M(x0 , x0 ) + U(w(x) − x0 ) − U(w(x0 ) − x0 )
(7.68) = M(x, x0 ).
Similarly, for x0 ≤ x, using the inequality in (7.64),
Z x
(7.69) M(x, x) = M(x , x ) +
0 0
N0 (z)dz
x0
Z x
(7.70) ≥ M(x , x ) +
0 0
U0 (w(z) − z)w0 (z)dz
x0
Z x
(7.71) ≥ M(x , x ) +
0 0
U0 (w(z) − x0 )w0 (z)dz
x0
(7.72) = M(x , x ) + U(w(x) − x0 ) − U(w(x0 ) − x0 )
0 0
(7.73) = M(x, x0 ).
Therefore M(x, x) ≥ M(x, x0 ) for all x0 , so that the family at x behaves optimally by bequeathing
x.
Now consider a family located at x ∈ [0, θ]. We know that its total wealth equals Ω(w∗ , r) =
w(θ) ≤ w(x) for all x ≥ θ, so by a standard single-crossing argument, and the observations of
the previous paragraph, that family will never bequeath more than θ. Therefore this family
must behave just as in a Becker-Tomes benchmark world with prices (w∗ , r), so its assigned
bequest in optimal.
Proof of Proposition 7.11: Suppose, on the contrary, that there are two steady state wage
functions (modulo equivalent representations). Denote these by w1 and w2 , and observe
from Proposition 7.9 that each of them must belong to the two-phase family. Let w1 and w2
be the two baseline wages, and let θi = θ(wi ), for i = 1, 2. Without loss of generality suppose
that θ2 ≥ θ1 .
These two wage functions must cross, otherwise if the profit-maximization (support)
condition is satisfied at one of them it will not be satisfied at the other. Beyond θ2 both
wage functions satisfy the same differential equation (7.17), which rules out a crossing in
this region. The functions also cannot cross below θ1 since both wage functions are r-linear
in this region. So θ1 < θ2 , and the functions cross at some x̄ ∈ (θ1 , θ2 ).
Now proceed as in the proof of Lemma 7.18. Pick some x in (x̄, θ2 ), inhabited under w2 ; then
pick x0 ∈ (x̄, x) such that w1 (x0 ) = w2 (x) > 0.
By construction, Ω(w2 , r) > w2 (x) (x is on the r-linear segment of w2 ), and so W̃(w2 (x); w2 , r) >
w2 (x). By Lemma 7.5,
(1 + r)V 0 (w2 (x)) > U0 (w2 (x) − x).
Using w1 (x0 ) = w2 (x), x0 < x and the concavity of U, we must conclude that
(1 + r)V 0 (w1 (x0 )) > U0 (w1 (x0 ) − x0 ),
but this contradicts the fact that w1 satisfies (7.17) at x0 .
Proof of Observation 7.6. Condition P tells us that for some two-phase wage function w,
c(r, w) ≤ 1. Define a new wage function ŵ that is r-linear from the same baseline wage as
that for w; then by Lemma 7.15, ŵ(x) ≤ w(x) for all x. It follows that c(r, ŵ) ≤ 1 as well. The
existence of the required baseline wage a now follows from the same argument used in the
proof of Proposition 7.10.
Proof of Proposition 7.12. First assume that (7.19) fails; using the same technique as in the
proof of Proposition 7.10, it is easy to see that the r-linear wage function starting at a is an
equal steady state. Given Proposition 7.11, this completes the proof.
Indeed, by the characterization result of proposition 7.9, an equal steady state wage function
must be the wage function that starts at a. If, therefore, (7.19) holds, that proposition assures
us that an equal steady state cannot exist.
7.4.11 Informal Observations on Equilibrium Dynamics. [This section needs rewriting

to harmonize notation and cross-references with the material discussed earlier. What follows
has been extracted from Mookherjee and Ray (AER 2002).]
We now turn to the question of non-steady state dynamics. The basic ideas are simple,
though the details are complicated and thus suppressed below. A distribution of wealth
(past bequests plus current income) prevails at any date; this will map into a distribution
of wealth for the next generation. Of course, the analysis will be different depending on
whether condition (??) holds or fails, and in addition there are several different kinds of
initial conditions to consider. We report on a single case, but one that holds particular
interest. We assume that initial wealth is perfectly equally distributed, yet we suppose that a
steady state is incompatible with perfect equality; i.e. that (??) holds.
Begin, then, with a single wealth level which we shall call W0 , commonly held by every
member of generation 0. The key to understanding the dynamics is the following simple
but powerful observation.
Observation 7.7. Suppose that next period’s wage function is given by w1 , with lowest wage
w1 . Let W1 ≡ Ψ(W0 , w1 ). Then for every x such that w1 (x) ≤ W1 , it must be the case that
¯ (x) = w + (1 + r)x. ¯
w1
¯1
This follows from similar arguments made earlier: the rate of return on educational
investment must equal r over the entire range spanned by financial bequests for any given
occupation. Observation 7.7 has an interesting corollary.
Observation 7.8. For any N, there exists a threshold such that starting from any equal initial wealth
above this threshold, there is perfect equality for at least N generations.
To see this, recall the wage function w∗ constructed earlier . For any N, define a threshold
— call it W̃ — such that if W0 ≥ W̃, then Ψ(N) (W0 , w∗ ) ≥ w̄∗ , where Ψ(N) is the function Ψ
¯
w(x), W
W0
W1
WN-1
WN
WN+1
w = w* up to N-1
w at N
w
_*
w
_ w at N+1
x
xN+1 xN M
Figure 7.8. Symmetry-Breaking along Equilibrium Paths.
iterated N times. In words, we choose an initial wealth high enough so that N iterations
will not suffice to bring wealth down to w̄∗ , in the pure bequest model with baseline wage
w∗ . If the wage function for the next N periods is given by w∗ , this is perfectly consistent
¯
with equilibrium behavior for those N periods. All occupations are equally valuable, so
individuals are indifferent over these choices. Moreover, there will be full equality over this
epoch.
Observation 7.7 implies that this is the only possible outcome for the first N periods. Any
other candidate wage function would have to exhibit a return of r over occupations in
which the wage fell short of wealth. This implies right away that no other equilibrium
wage function can exist with the same starting wage w∗ . Using Observation 7.7 again and
employing an argument similar to that used to establish ¯ uniqueness, we may also rule out
any candidate wage function with a different starting wage than w∗ .
¯
So perfect equality can prevail for a substantial number of periods. But — if (??) holds —
it cannot prevail forever. Because Ω(w∗ ) < w̄∗ , there must eventually come a date T when
¯
the recursion ψ(T) (W0 , w∗ ) dips below w̄∗ . At this stage, the symmetry of equilibrium is
broken and the economy ¯ must depart from the wage function w∗ . If not, all occupations
with wages that exceed ψ(T) (W0 , w∗ ) would remain unoccupied. This cannot be: the wages
¯
of such occupations must rise, yielding rates of return that exceed their counterpart on
financial bequests. For all other occupations, Observation 7.7 is applicable in full force, and
the occupational rate of return remains anchored to the financial rate.
Figure 11.3 describes the resulting “distortion” in the wage function. The diagram displays
N initial periods in which the wage function w∗ prevails, and there is no inequality. At date
N, wealth dips into the support of w∗ , and no occupation more expensive than xN can be
supported under the old rates of return. The economy must react by lowering the wages of
all occupations below xN , and by raising the rates of return to all other occupations. These
higher rates encourage the acquisition of such professions, and in so doing must generate
inequality. Of course, all generation-N individuals are still indifferent between the various
choices. But this is the last generation for which all payoffs are equal. The higher rates of
return on the more expensive occupations must inevitably result in inequality among the
next generation.
Once such inequality sets in, it will not go away. Familiar single-crossing arguments
guarantee that descendants of higher-wealth individuals must occupy the richer professions
and receive higher payoffs. The subsequent dynamics are complicated by the fact that the
wealth distribution is no longer degenerate. In the special case where there are just two
occupations (involving different training costs), it can be shown that the economy converges
to an unequal steady state, with inequality rising over time. Moreover, the rise in inequality
is augmented in the presence of financial bequests: it turns out that individuals in the
unskilled occupation at any date are less wealthy than their parents, and make smaller
financial bequests to their children than they received.
In the continuum case, the wealth distribution at any date must retain a mass point at its
lower bound, and it is this mass point which will be spread over all the “low” occupations,
which continue to bear the financial rate of return r. In the relatively “high” occupations, the
rate of return will continue its departure from r, attracting more individuals into this zone.
Figure 11.3 illustrates this with yet another iteration for date N + 1. The new wage function is
given by the dashed curve. It must be linear (with slope 1 + r) up to the new threshold xN+1 ,
and then rises even more steeply than before, intersecting both the previous wage functions
from below. This rise induces some fresh symmetry-breaking, as new dynasties from the
“cheap” occupations seek the higher rates of return. If this process converges (a subject of
our current research), it must be to precisely the steady state we have described earlier. At
this steady state, there is a mass point of individuals with identical wealth, and among such
individuals there is a simple tradeoff between occupational choice and financial bequests,
among which they are indifferent. But there will also be a positive measure of individuals
arrayed over varying levels of wealth (and utility).
CHAPTER 8
Polarization and Conflict
As the struggle proceeds, the whole society breaks up more and more into
two hostile camps, two great, directly antagonistic classes: bourgeoisie
and proletariat. The classes polarize, so that they become internally more
homogeneous and more and more sharply distinguished from one another
in wealth and power. (Morton Deutsch, 1971)
8.1 Introduction
Conflict is a fundamental part of life in many developing countries. This section of the
course will try to set up models of conflict and also study its correlates.
Some basic facts: between 1945–1999, battle deaths in 25 interstate wars were approximately
3.33m. Compared to that, deaths in civil wars have totalled around 16.2m. Such wars have
been experienced in 73 states, with a median duration of six years. In 1999, there were 25
ongoing conflicts. It is reasonable to say that matters have only become worse in the twenty
first century.
These numbers are symptomatic of ongoing unrest, violence around the world. They
obviously have big effects for GDP and growth: one writer places the narrow economic
costs of conflict at around 8% of GDP.
By conflict we don’t just mean violent conflict. Demonstrations, strikes, processions: all of
these also cause a substantial amount of economic waste.
One sort of conflict may simply come from class antagonisms: the presence of rich and
poor subclasses in the population. This is classic redistributive conflict where redistribution
occurs, or is sought, from rich to poor. But it has also been argued that “ethnic divisions”
— broadly defined — may be a significant determinant of conflict. For instance, writers
such as Samuel Huntington (1996) have argued for a cultural perspective on conflict and
war. After all, there is no particular reason why redistribution must occur across economic
classes. It could happen between religious, ethnic, or spatially separated groups. Resources
are resources, no matter where they come from.
134 Polarization and Conflict
The monumental threatise on ethnic conflict is by Donald Horowitz (1985), who summarizes
thus:
“The Marxian concept of class as an inherited and determinative affiliation

finds no support in [the] data. Marx’s conception applies with far less
distortion to ethnic groups. Ethnic membership us generally given at birth,
. . . and have considerable power to generate conflict . . .
In much of Asia and Africa, it is only modest hyperbole to assert that
the Marxian prophecy has had an ethnic fulfillment.”
As I’ve said above, this doesn’t mean that ethnicity is necessarily a “primordial” or intrinsic
source of conflict. Ethnicity may simply be a relevant marker for carving out a larger
share of a smaller pie. Very often, ethnic groups are not particularly ranked by income:
Sinhalese/Tamils in Sri Lanka, Malays/Chinese in Malaysia, Hausa/Yoruba in Nigeria,
Serbs/Muslims in Bosnia, Dalits/low-income-Muslims in Gujarat, Basques/others in Spain.
In other situations, marked socio-economic differences do exist: Hutu/Tutsi in Rwanda
and Burundi, caste divisions in Bihar, Bengalis in pre-Bangladesh Pakistan, Muslims in the
Philippines, Kurds in Iraq, Sikhs in the Punjab, or the Naxalites in Eastern and SouthEastern
India.
Some interesting questions:
What sort of distributions make for conflict?
Distributions over what ? Ethnic groups or income classes? The answer would presumably
depend on what the salient marker is for conflict.
How might we decide whether unranked ethnicity or ranked class is salient in conflict?
8.2 What Creates Conflict?
8.2.1 Variables. Much (though not all) of the empirical literature on conflict throws in the
kitchen sink to see what washes up. On the left hand side goes in one of several alternative
dependent variables: many measures of “conflict”, ranging all the way from incidence of
demonstrations, procesions, or strikes, through riots and on to civil war.
Even with a specific choice such as civil war (used here) we need defining criteria (see Singer-
Small (1982), Licklider (1993), Doyle-Sambanis (2000), Fearon-Laitin (2003)). One important
distinction to be made is between onset and incidence. Did the conflict start in that year of
measurement, or was it ongoing? As we will see later this makes a big difference. Another
important question is the minimum number of deaths before a conflict is classifiable as such:
for civil wars, this is a matter of ongoing debate.
On the right hand-side goes in a slew of explanatory variables:
Economic: per-capita income, inequality of income or wealth, resource holdings . . .
Geographical: mountainous terrain, separation from capital city . . .
Polarization and Conflict 135
Political: “extent of democracy”, prior war . . .

And, of course
Ethnic: Measures of ethnolinguistic diversity drawn from the World Christian Encyclopedia,
Encyclopedia Britannica, Atlas Narodov Mira, CIA FactBook. Measures of religious diversity
drawn from L’Etat des Religions dans le Monde, World Christian Encyclopedia, The
Statesman’s Yearbook.
This last information is typically put into an index of diversity. An index with particularly
wide currency is the ethnolinguistic fractionalization index, or ELF (of course, there’s nothing
to stop its use for religious diversity as well).
To define this, say there are M groups. n j is the population share of group j. Then
M
X
E= n j (1 − n j ).
j=1
This has an obvious relationship to the Gini index of inequality if we take inter-group
distance to be equal to one.
ELF is widely used in empirical work (see Taylor and Hudson (1972), Mauro(1995), Easterly
and Levine (1997), Alesina et al. (2003), Vigdor (2002), Collier and Hoeffler (2002), Fearon
and Laiton (2003), and many others).
8.2.2 Some Findings. Perhaps the most robust finding in many different empirical studies
is that per-capita income is negatively related to conflict. A negative and significant
association is to be observed in the cross-sectional studies carried out by Collier and Hoeffler
(2002), Fearon and Laitin (2003), Montalvo and Reynal-Querol (2005) and several others. As
Fearon and Laitin (2003) summarize,
“Per capita income (measured as thousands of 1985 U.S. dollars and lagged
one year) is strongly significant in both a statistical and a substantive sense
...”
It hardly needs mentioning that no causal inference can be drawn from such an association.
Relatively speaking, the best study that addresses this question is Miguel, Satyanath
and Sergenti (2004), which instruments for GDP growth shocks using rainfall. Such an
instrunment would be unacceptably weak in many of the richer developing countries, so
they restrict attention to a sample of 41 sub-Saharan African countries over the period
1981-1999. Their main results are best summarized in their own words:
“Using the comprehensive new database of conflicts developed by the

International Peace Research Institute of Oslo, Norway, and the University
of Uppsala, Sweden, we find that GDP growth is significantly negatively
related to the incidence of civil conflict in sub-Saharan Africa during the
period 1981–99 across a range of regression specifications, including some
with country fixed effects. The relationship between GDP growth and the
incidence of civil wars is extremely strong: a five-percentage-point drop
in annual economic growth increases the likelihood of a civil conflict (at

least 25 deaths per year) in the following year by over 12 percentage points
— which amounts to an increase of more than one-half in the likelihood
of civil war. Other variables that have gained prominence in the recent
literature — per capita GDP level, democracy, ethnic diversity, and oil
exporter status — do not display a similarly robust relationship with the
incidence of civil wars in sub-Saharan Africa. In the second main result,
we find — perhaps surprisingly — that the impact of income shocks on
civil conflict is not significantly different in richer, more democratic, more
ethnically diverse, or more mountainous African countries or in countries
with a range of different political institutional characteristics.”
To be sure, the exclusion restriction behind the use of rainfall as an instrument could be
violated if rainfall has some direct effect on conflict via some other pathway. For instance,
floods could destroy the road network and make it harder to contain government troops
— more conflict (though this is a bias that runs the other way from the instrumental
direction, which may actually strengthen the results). But of course, one could cook up
other connections, such as floods make it harder for government troops and rebels to engage
each other, which reduces conflict (in terms of deaths). The authors attempt to take care of
some of these effects in the paper.
What lies behind the connection behind poverty and conflict? What comes to mind most
immediately is the fact that poverty reduces the opportunity cost of conflict labor, making
it easier to use labor for some other activity. Of course, with poverty all around there may
not be much to appropriate, so this suggests that it might be more interesting to examine the
effects of (growth shocks)*inequality.
Some evidence that opportunity costs lie at the heart of the story comes from the paper
by Dube and Vargas (2007), which studies the effect of coffee and oil prices on conflict in
Colombia. Oil prices affect government revenues which are there to be grabbed, so one
would imagine that higher prices for oil positively affect conflict. In contrast, if coffee-
growing is an economic activity that competes with resources given over to conflict (such
as labor!), then an increase in coffee prices would reduce conflict. Thus if the opportunity
cost story is right, one would expect that oil and coffee have opposite effects on conflict, and
indeed this is what Dube and Vargas find.
The opportunity cost argument is explicitly made by Collier and Hoeffler (1998), though
it is fairly obvious and appears implicitly in many papers. In contrast, Fearon and laitin
(2003) have argued that poverty is likely correlated with low government capabilities and
infrastructure that makes it harder to crack down on insurgencies. In support of this line
of reasoning they show that “mountainous terrain” is significant in “explaining” conflict,
showing thereby that infrastructure and geography matter. This line of argument requires
more investigation, however.
A second important finding is that the inequality of income appears to have at best an
ambiguous effect on conflict. Some early papers on this subject are Nagel (1974) for Vietnam,
Midlarski (1988) and Muller, Seligson and Fu (1989) (the last two are on land inequalities).
Specifically, under several measures of inequality such as the Gini, conflict appears to be
low both for low and for high values of inequality. This is a theme to which we will return
in more detail below.
As a related matter, we’ve already noted that measures of ethnic dispersion are closely
related to “inequality” — i.e., diversity — across ethnic groups. These, too, fail to matter.
The same findings appear and reappear in Collier and Hoeffler (1998, 2002), Fearon and
Laitin (2003), as well as Miguel et al (2004). Fearon and Laitin (2003) conclude thus:
“The estimates for the effect of ethnic and religious fractionalization are
substantively and statistically insignificant . . . The empirical pattern is thus
inconsistent with . . . the common expectation that ethnic diversity is a
major and direct cause of civil violence.”
This isn’t to say that ethnic or religious fractionalization cannot indirectly affect conflict. It
might, via reduced GDP (Alesina et al. (2003)), reduced GDP growth (Easterly and Levine
(1997)), or poor governance (Mauro (1995)). The claim really is that there is no direct effect.
An important theme that we pursue below is the development of a measure of “polarization”,
which we show is different from inequality. Is a Gini-like index like ELF a good measure?
Listen to Horowitz again:
“I have intimated at various points that a system with only two ethnic
parties . . . is especially conflict prone . . . In dispersed systems, group
loyalties are parochial, and ethnic conflict is localized; it ‘could put one
of a series of watertight compartments out of order, but it could not make
the ship of state sink . . . ’ The demands of one group can sometimes be
granted without injuring the interests of others . . . ”
On the other hand, continues Horowitz,
“A centrally focused system [with few groupings] possesses fewer cleav-

ages than a dispersed system, but those it possesses run through the whole
society and are of greater magnitude. When conflict occurs, the center has
little latitude to placate some groups without antagonizing others.”
The conflictual power of broad cleavages is of course an older theme: read the quotation at
the beginning of this chapter!
8.3 The Identity-Alienation Framework
Following Esteban and Ray (1991, 1994) and Duclos, Esteban and Ray (2004), I develop
below the identification-alienation framework. The idea is simple: polarization is related to
the alienation that individuals and groups feel from one another, but such alienation is fuelled
by notions of within-group identity. In concentrating on such phenomena, I do not mean to
suggest that instances in which a single isolated individual runs amok with a machine gun
are rare, or that they are unimportant in the larger scheme of things. It is just that these are
not the objects of our enquiry. We are interested in the correlates of organized, large-scale
social unrest — strikes, demonstrations, processions, widespread violence, and revolt or
rebellion. Such phenomena thrive on differences, to be sure. But they cannot exist without
notions of group identity either.
This brief discussion immediately suggests that inequality, inasmuch as it concerns itself with
interpersonal alienation, captures but one aspect of polarization. To be sure, there are some
obvious changes that would be branded as both inequality- and polarization-enhancing.
For instance, if two income groups are further separated by increasing economic distance,
inequality and polarization would presumably both increase. However, local equalizations
of income differences at two different ranges of the income distribution will most likely lead
to two better-defined groups — each with a clearer sense of itself and the other. In this case,
inequality will have come down but polarization may be on the rise.
Imagine, then, that society is divided into “groups” (economic, social, religious, spatial...)
Identity. There is “homogeneity” within each group.
Alienation. There is “heterogeneity” across groups.
The IA framework presumes that such a situation is inherently conflictual. In the words of
Esteban and Ray (1994),
“We begin with the obvious question: why are we interested in polariza-
tion? It is our contention that the phenomenon of polarization is closely
linked to the generation of tensions, to the possibilities of articulated
rebellion and revolt, and to the existence of social unrest in general . . . At
the same time, measured inequality in such a society may be low.”
Does the standard theory of inequality measurement fit? Recall the
Pigou-Dalton Transfers Principle. A transfer of resources from a relatively poor to a relatively

rich individual must raise income inequality.
This principle forms the building block for all measures of inequality. But now look at this
example. Start here:
Density
Income or Wealth
and go here:
Density
Income or Wealth
By all measures of inequality that are consistent with the Lorenz criterion (or equivalently,
second-order stochastic domination), inequality has come down. Yet there is something
disturbing about that: social tensions could be going up as two well-defined and distinct
groups begin to form. There could be situations here in which polarization is going up.
Of course, this isn’t to suggest that polarization — whatever it is — is always different from
inequality. If there is a “global compression” of the distribution, we would expect both
inequality and polarization to fall. See these diagrams. Start here:
Density
Income or Wealth
and go here:
Density
Income or Wealth
But these two examples also tell us something else. It tells us that polarization may not be a
“local construct”. In both the examples, we have a compression taking place, yet they have
very different implications. A “local move” may have different effects depending on the
overall distribution. In contrast, inequality — as captured by the Pigou-Dalton principle —
is a local construct.
Observe, too, that the notion of “groups” may be quite general:
Economic: income- or wealth-based (class)
Social: religious, linguistic, geographical, political groupings.

But a natural notion of “distance” across economic groups makes income-based polarization
an easier starting point.
8.4 A Measure of Polarization
The task of this section is to develop axiomatically a measure of polarization. Our “inputs”
are various distributions of income or wealth on different populations (more precisely, density
functions with varying populations).
Our “output” is a measure of polarization for each distribution.
As we’ve tried to motivate earlier, each individual feels two things:
Identification with people of “similar” income.
[Use as proxy the height of density f (x) at income x.]
Alienation from people with “dissimilar” income.
[Income distance |y − x| of y from x.]

We therefore describe the effective antagonism of x towards y as a function T(i, a), where i is
the identification that a person at income x feels, and a is the income distance between x
and y. We suppose that T(0, a) = T(i, 0) = 0: both a sense of identification and alienation is
needed to fuel an effective sense of antagonism.
We view polarization as the “sum” of all such antagonisms over the population:
Z Z
P( f ) =

T f (x), |x − y| f (x) f (y)dxdy
This is not very useful as it stands. Way too much depends on the choice of the function T.
But hopefully it is a good starting point.
The axioms we use are based on densities that are unions of one or more basic densities. These
are symmetric, unimodal density functions f with compact support.
[By symmetry we mean that f (m − x) = f (m + x) for all x ∈ [0, 1], and by unimodality we
mean that f is nondecreasing on [0, m].]
A basic density (or indeed any density)
can be population-scaled: g(y) = p f (y).
can be income-scaled: g(y) = (1/µ) f (x/µ).
can undergo a slide to the right or left: g(y) = f (y − x), and
can be squeezed.
Take a closer look at this last one. Let f be any density and let λ lie in (0, 1]. A λ-squeeze of f
is a transformation f λ :
1 x − [1 − λ]m

(8.1) f λ (x) ≡ f .
λ λ
Scalings, slides and squeezes partition the space of all densities. Each element of the partition
can be associated with a root, a basic density on [0, 2] with mean 1.
Axiom 1. If a distribution is just a single basic density, a “global compression” of that density
cannot increase polarization.
That is: global compression (the move described in the diagram below) cannot raise
polarization.
Income or Wealth
Axiom 2. If a symmetric distribution is composed of three disjoint scalings of the same basic
density, then a compression of the side densities cannot reduce polarization.
This is illustrated in the diagram below.

Density
Income
Axiom 3. Consider a symmetric distribution composed of four basic densities drawn from
the same root. Slide the two middle densities to the side as shown. Then polarization must
go up.
This is illustrated in the diagram below:

Density
Income
Our last axiom states that polarization rankings must be invariant to population scalings.
Axiom 4. [Population Neutrality.] Polarization comparisons are unchanged if both
populations are scaled up or down by the same percentage.
We can now state:
Theorem 8.1. A polarization measure satisfies Axioms 1–4 if and only if it is proportional to
Z Z
(8.2) f (x)1+α f (y)|y − x|dydx,
where α lies between 0.25 and 1.
8.5 Proof of Theorem 8.1
Note: this section is technical (though illustrative), and should probably be omitted unless
you are comfortable with long and detailed arguments.
First, we show that axioms 1–4 imply (8.2). The lemma below follows from Jensen’s
inequality; proof omitted.
Lemma 8.1. Let g be a continuous real-valued function defined on IR such that for all x > 0 and all δ
with 0 < δ < x,
Z x+δ
1
(8.3) g(x) ≥ g(y)dy.
2δ x−δ
Then g must be a concave function.
In what follows, remember that our measure only considers income differences across people,
so that we may slide any distribution to left or right as we please.
Lemma 8.2. The function T must be concave in a for every i > 0.
Proof. Fix x > 0, some i > 0, and δ ∈ (0, x). Consider three basic densities as in Axiom 2 but
specialize as shown in Figure 8.1; each is a transform of a uniform basic density. The bases
h/λ h/λ
i
h h
2λδ 2λδ
2δ 2ε 2δ
-x 0 x
Figure 8.1
are centered at −x, 0 and x. The side densities are of width 2δ and height h, and the middle
density is of width 2 and height i. We shall vary and h but to make sure that Axiom 2
applies, we choose > 0 such that δ+ < x. A λ-squeeze of the side densities simply contracts
their base width to 2λδ, while the height is raised to h/λ. For each λ, decompose the measure
(??) into five components. (a) The “internal polarization” Pm of the middle rectangle. This
component doesn’t vary with λ so there will be no need to explicitly calculate it. (b) The
“internal polarization” Ps of each side rectangle. (c) Total effective antagonism, Ams felt by
inhabitants of the middle towards each side density. (d) Total effective antagonism Asm felt
by inhabitants of each side towards the middle. (e) Total effective antagonism Ass felt by
inhabitants of one side towards the other side. Each of these last four terms appear twice,
so that (writing everything as a function of λ),
(8.4) P(λ) = Pm + 2Ps (λ) + 2Ams (λ) + 2Asm (λ) + 2Ass (λ),
Now we compute the terms on the right hand side of (8.4). First,
Z x+λδ Z x+λδ
1
Ps (λ) = 2 T(h/λ, |b0 − b|)h2 db0 db,
λ x−λδ x−λδ
where (here and in all subsequent cases) b will stand for the “origin” income (to which the
identification is applied) and b0 the “destination income” (towards which the antagonism is
felt). Next,
1 x+λδ
Z Z
Ams (λ) = T(i, b0 − b)ihdb0 db.
λ − x−λδ
Third,
1 x+λδ
Z Z
Asm (λ) = T(h/λ.b − b0 )hidb0 db,
λ x−λδ −
And finally,
Z −x+λδ Z x+λδ
1
Ass (λ) = 2 T(h/λ, b0 − b)h2 db0 db.
λ −x−λδ x−λδ
The axiom requires that P(λ) ≥ P(1). Equivalently, we require that [P(λ) − P(1)]/2h ≥ 0 for
all h, which implies in particular that
P(λ) − P(1)
(8.5) lim inf ≥ 0.
h→0 2h
If we divide through by h in the individual components calculated above and then send h to
0, it is easy to see that the only term that remains is Ams . Formally, (8.5) and the calculations
above must jointly imply that
1 x+λδ
Z Z Z Z x+δ
0 0
(8.6) T(i, b − b)db db ≥ T(i, b0 − b)db0 db,
λ − x−λδ − x−δ
and this must be true for all λ ∈ (0, 1) as well as all ∈ (0, x − δ). Therefore we may insist
on the inequality in (8.6) holding as λ → 0. Performing the necessary calculations, we may
conclude that
1 1 x+δ
Z Z Z
(8.7) T(i, x − b)db ≥ T(i, b0 − b)db0 db
− − x−δ
for every ∈ (0, x − δ). Finally, take to zero in (8.7). This allows us to deduce that
Z x+δ
(8.8) T(i, x) ≥ T(i, b0 )db0 .
x−δ
As (8.8) must hold for every x > 0 and every δ ∈ (0, x), we may invoke Lemma 8.1 to conclude
that T is concave in x for every i > 0.
Q.E.D.
Lemma 8.3. Let g be a concave, continuous function on IR+ , with g(0) = 0. Suppose that for each a
and a0 with a > a0 > 0, there exists ∆
¯ > 0 such that
(8.9) g(a + ∆) − g(a) ≥ g(a0 ) − g(a0 − ∆)

for all ∆ ∈ (0, ∆).
¯ Then g must be linear.
The proof is straightforward and is omitted.

Lemma 8.4. There is a continuous function φ(i) such that T(i, a) = φ(i)a for all i and a.
Proof. Fix a and a0 with a > a0 > 0, and i > 0. Consider four basic densities as in Axiom
3 (see Figure ??) but specialize as shown in Figure 8.2; each is a transform of a uniform
basic density. The bases are centered at locations −y, −x, x and y, where x ≡ (a − a0 )/2
and y ≡ (a + a0 )/2. The “inner” densities are of width 2δ and height h, and the “outer”
densities are of width 2 and height i. We shall vary different parameters (particularly x) but
to ensure disjoint support we assume throughout that < x and δ + < y − x − ∆ ¯ for some
∆¯ > 0. Again, decompose the polarization measure (??) into several distinct components.
(a) The “internal polarization” of each rectangle j; call it P j , j = 1, 2, 3, 4. These components
are unchanged as we change x so there will be no need to calculate them explicitly. (b)
h h
i i
1 2 3 4
2ε 2δ 2δ 2ε
-y -x x y
Figure 8.2
Total effective antagonism A jk (x) felt by inhabitants of rectangle j towards rectangle k (we
emphasize dependence on the parameter x). Thus total polarization P(x) is given by
4
X XX
P(x) = Pj + A jk (x)
j=1 j k6= j
4
X
= P j + 2A12 (x) + 2A13 (x) + 2A21 (x) + 2A31 (x) + 2A23 (x) + 2A14 ,
j=1
where the second equality simply exploits obvious symmetries and A14 is noted to be
independent of x. Let’s compute the terms in this formula that do change with x. We have
Z −y+ Z −x+δ
A12 (x) = T(i, b0 − b)ihdb0 db,
−y− −x−δ
Z −y+ Z x+δ
A13 (x) = T(i, b0 − b)ihdb0 db,
−y− x−δ
Z −x+δ Z −y+
A21 (x) = T(h, b − b0 )ihdb0 db,
−x−δ −y−
Z x+δ Z −y+
A31 (x) = T(h, b − b0 )ihdb0 db,
x−δ −y−
and
Z −x+δ Z x+δ
A23 (x) = T(h, b − b0 )h2 db0 db.
−x−δ x−δ
Now, the axiom requires that P(x + ∆) − P(x) ≥ 0. Equivalently, we require that [P(x + ∆) −
P(1)]/2ih ≥ 0 for all h, which implies in particular that
P(x + ∆) − P(x)
lim inf ≥ 0.
h→0 2ih
Using this information along with the computations for P(x) and the various A jk (x)’s, we see
(after some substitution of variables and transposition of terms) that
Z −y+ Z x+δ
[T(i, b0 − b + ∆) − T(i, b0 − b)] db0 db
−y− x−δ
Z −y+ Z −x+δ
≥ [T(i, b0 − b) − T(i, b0 − b − ∆)] db0 db,
−y− −x−δ
Dividing through by δ in this expression and then taking δ to zero, we may conclude that
Z −y+ Z −y+
[T(i, x − b + ∆) − T(i, x − b)] db ≥ [T(i, −x − b) − T(i, −x − b − ∆)] db,
−y− −y−
and dividing this inequality, in turn, by and taking to zero, we see that
T(i, a + ∆) − T(i, a) ≥ T(i, a0 ) − T(i, a0 − ∆),
where we use the observations that x + y = a and y − x = a0 . Therefore the conditions of
Lemma 8.3 are satisfied, and T(i, .) must be linear for every i > 0 since T(0, a) = 0. That is,
there is a function φ(i) such that T(i, a) = φ(i)a for every i and a. Given that T is continuous
by assumption, the same must be true of φ.
Q.E.D.
Lemma 8.5. φ(i) must be of the form Kiα , for constants (K, α) 0.
Proof. As a preliminary step, observe that

(8.10) φ(i) > 0 whenever i > 0.
otherwise Axiom 3 would fail for configurations constructed from rectangular basic densities
of equal height i. We first prove that φ satisfies the fundamental Cauchy equation
(8.11) φ(p)φ(p0 ) = φ(pp0 )φ(1)
for every (p, p0 ) 0. To this end, fix p and p0 and define r ≡ pp0 . In what follows, we assume
that p ≥ r.1 Consider a configuration with two basic densities, both of width 2, the first
centered at 0 and the second centered at 1. The heights are p and h (where h > 0 but soon to
be made arbitrarily small). A little computation shows that polarization in this case is given
by
Z Z 1+
P = ph[φ(p) + φ(h)]{ (b0 − b)db0 db}
− 1−
Z Z
+[p φ(p) + h φ(h)]{
2 2
|b0 − b|db0 db}
− −
83 2
(8.12) = 42 ph[φ(p) + φ(h)] + [p φ(p) + h2 φ(h)],
3
where the first equality invokes Lemma 8.4. Now change the height of the first rectangle to
r. Using (8.10) and p ≥ r, it is easy to see that for each , there exists a (unique) height h()
1If r ≥ p, simply permute p and r in the argument below.
for the second rectangle such that the polarizations of the two configurations are equated.
Invoking (8.12), h() is such that
2
ph[φ(p) + φ(h)] + [p2 φ(p) + h2 φ(h)]
3
2
(8.13) = rh()[φ(r) + φ(h())] + [r2 φ(r) + h()2 φ(h())].
3
By Axiom 4, it follows that for all λ > 0,
2
λ2 ph[φ(λp) + φ(λh)] + [(λp)2 φ(λp) + (λh)2 φ(λh)]
3
2
(8.14) = λ rh()[φ(λr) + φ(λh())] + [(λr)2 φ(λr) + [λh()]2 φ(λh())].
2
3
Notice that as ↓ 0, h() lies in some bounded set. We may therefore extract a convergent
subsequence with limit h0 as ↓ 0. By the continuity of φ, we may pass to the limit in (8.13)
and (8.14) to conclude that
(8.15) ph[φ(p) + φ(h)] = rh0 [φ(r) + φ(h0 )]
and
(8.16) λ2 ph[φ(λp) + φ(λh)] = λ2 rh0 [φ(λr) + φ(λh0 )].
Combining (8.15) and (8.16), we see that
φ(p) + φ(h) φ(r) + φ(h0 )
(8.17) = .
φ(λp) + φ(λh) φ(λr) + φ(λh0 )
Taking limits in (8.17) as h → 0 and noting that h0 → 0 as a result (examine (8.15) to confirm
this), we have for all λ > 0,
φ(p) φ(r)
(8.18) = .
φ(λp) φ(λr)
Put λ = 1/p and recall that r = pp0 . Then (8.18) yields the required Cauchy equation (8.11).
To complete the proof, recall that φ is continuous and that (8.10) holds. The class of solutions
to (8.11) (that satisfy these additional qualifications) is completely described by φ(p) = Kpα
for constants (K, α) 0 (see, e.g., Aczél [1966, p. 41, Theorem 3]).
Lemmas 8.4 and 8.5 together establish “necessity”, though it still remains to establish the
bounds on α. We shall do so along with our proof of “sufficiency”, which we begin now.
Lemma 8.6. Let f be a basic density with mass p and mean µ on support [a, b]. Let m ≡ µ − a and
let f ∗ denote the root of f . Then, if f λ denotes some λ-squeeze of f ,
Z 1 (Z 1 Z 1 )
λ
(8.19) P(F ) = 4kp (mλ)
2+α 1−α ∗
f (x)1+α
f (y)(1 − y)dy +
∗ ∗
f (y)(y − x)dy dx
0 0 x
for some constant k > 0.
Proof. Recall that a slide of f has no effect on the computations, so we may as well set a = 0
and b = 2m, where m = µ − a is now to be interpreted as the mean. Given (8.2),
Z Z
(8.20) P(F) = k f (x)1+α f (y)|y − x|dydx
for some k > 0. Using the fact that f is symmetric, we can write
Z m Z 2m
P(F) = 2k f (x0 )1+α f (y0 )|x0 − y0 |dy0 dx0
0 0
Z m (Z x0 Z m
= 2k f (x )0 1+α
f (y )(x − y )dy +
0 0 0 0
f (y0 )(y0 − x0 )dy0
0 0 x0
Z 2m )
(8.21) + 0 0
f (y )(y − x )dy }dx 0 0 0
.
m
Examine the very last term in (8.21). Change variables by setting z ≡ 2m − y0 , and use
symmetry to deduce that
Z 2m Z m
f (y0 )(y0 − x0 )dy0 = f (z)(2m − x0 − z)dz.
m 0
Substituting this in (8.21), and manipulating terms, we obtain
Z m (Z m Z m )
(8.22) P(F) = 4k 0 1+α
f (x ) f (y )(m − y )dy +
0 0 0
f (y )(y − x )dy dx0 .
0 0 0 0
0 0 x0
Now suppose that f λ is a λ-squeeze of f . Note that (8.22) holds just as readily for f λ as for
f . Therefore, using the expression for f given in (8.1), we see that
!1+α (Z m
y0 − (1 − λ)m
Z m
x0 − (1 − λ)m
!
λ
P(F ) = 4kλ −(2+α)
f f (m − y0 )dy0
(1−λ)m λ (1−λ)m λ
y0 − (1 − λ)m
Z m ! )
+ f (y − x )dy }dx .
0 0 0 0
x0 λ
x0 −(1−λ)m y0 −(1−λ)m
Perform the change of variables x00 = λ and y00 = λ . Then it is easy to see that
Z m (Z m Z m )
λ
P(F ) = 4kλ 1−α 00 1+α
f (x ) f (y )(m − y )dy +
00 00 00
f (y )(y − x )dy dx00 .
00 00 00 00
0 0 x00
To complete the proof, we must recover the root from f . To this end, first population-scale f∗
f to h, where h has mass 1. That is, f (z) = ph(z) for all z. Doing so, we see that
Z m (Z m Z m )
λ
P(F ) = 4kp λ
2+α 1−α 00 1+α
h(x ) h(y )(m − y )dy +
00 00 00
h(y )(y − x )dy dx00 .
00 00 00 00
0 0 x00
Finally, make the change of variables x = x00 /m and y = y /m. Noting
00 that f ∗ (z) = mh(mz),
we get (8.19).
Q.E.D.
Lemma 8.7. Let f and g be two basic densities with disjoint support, with their means separated by
distance d, and with population masses p and q respectively. Let f have mean µ on support [a, b]. Let
m ≡ µ − a and let f ∗ denote the root of f . Then for any λ -squeeze f λ of f ,
Z 1
(8.23) A( f λ , g) = 2kdp1+α q(mλ)−α f ∗ (x)1+α dx,
0
where A( f λ , g) denotes the total effective antagonism felt by members of f λ towards members of g.
Proof. Without loss of generality, let f have support [0, 2m] (with mean m) and g have
support [d, d + 2m] (where d ≥ 2m for disjoint supports). Using (8.20),
Z 2m Z d+2m 
A( f, g) = k 1+α 
 
f (x) 
 g(y)(y − x)dy dx
0 d
Z 2m
Z d+m Z d+2m 
= k f (x)1+α  g(y)(y − x)dy +
 
g(y)(y − x)dy dx

0 d d+m
Z 2m
Z d+m 
= k 1+α 
g(y)2(m + d − x)dy dx
 
f (x) 

0 d
Z 2m
= kq f (x)1+α (m + d − x)dx
0
Z m
= 2dkq f (x)1+α dx,
0
where the third equality exploits the symmetry of g,2 the fourth equality uses the fact that
R d+m
d
g(y) = q/2, and the final equality uses the symmetry of f .3 To be sure, this formula
applies to any λ -squeeze of f , so that
Z m
λ
A( f , g) = 2dkq f λ (x0 )1+α dx0
0
m !1+α
x0 − (1 − λ)m
Z
= 2dkqλ −(1+α)
f dx0 ,
(1−λ)m λ
x0 −(1−λ)m
and making the change of variables x00 = λ , we may conclude that
Z m
A( f λ , g) = 2dkqλ −α 00 1+α 00
f (x ) dx .
0
To complete the proof, we must recover the root f ∗ from f . As in the proof of Lemma 8.6,
first population-scale f to h, where h has mass 1. That is, f (z) = ph(z) for all z. Doing so, we
see that Z m
A( f λ , g) = 2dkp1+α qλ−α h(x00 )1+α dx00 .
0
Finally, make the change of variables x = x00 /m. Noting that f ∗ (z) = mh(mz), we get (8.23).
Lemma 8.8. Define, for any root f and α > 0,
R1
0
f (x)1+α dx
(8.24) ψ( f, α) ≡ R R .
1 1 R1
0
f (x)1+α 0
f (y)(1 − y)dy + x f (y)(y − x)dy dx
Then — for any α > 0 — ψ( f, α) attains its minimum value when f is the uniform root, and this
minimum value equals 3.
2That is, for each y ∈ [d, d + m], g(y) = g(d + 2m − (y − d)) = g(2d + 2m − y). Moreover, [y − x] + [(2d + 2m − y) − x] =
2(d + m − x).
3That is, for each x ∈ [0, m], f (x) = f (2m − x). Moreover, [m + d − x] + [m + d − (2m − x)] = 2d.
Proof. It will be useful to work with the inverse function

R1 R
1 R1
0
f (x) 1+α
0
f (y)(1 − y)dy + x
f (y)(y − x)dy dx
ζ( f, α) ≡ ψ( f, α)−1 = R1 .
f (x) 1+α dx
0
Note that ζ( f, α) may be viewed as a weighted average of
Z 1 Z 1
(8.25) L(x) ≡ f (y)(1 − y)dy + f (y)(y − x)dy
0 x
as this expression varies over x ∈ [0, 1], where the “weight” on a particular x is just
f (x)1+α
R1
0
f (z)1+α dz
which integrates over x to 1. Now observe that L(x) is decreasing in x. Moreover, by the
unimodality of a root, the weights must be nondecreasing in x. It follows that
Z 1
(8.26) ζ( f, α) ≤ L(x)dx.
0
Now
Z 1 Z 1
L(x) = f (y)(1 − y)dy + f (y)(y − x)dy
0 x
Z 1 Z x
= f (y)(1 − x)dy + f (y)(x − y)dy
0 0
Z x
1−x
(8.27) = + f (y)(x − y)dy.
2 0
Rx
Because f (x) is nondecreasing and integrates to 1/2 on [0, 1], it must be the case that 0 f (y)(x−
Rx
y)dy ≤ 0 (x − y)/2dy for all x ≤ 1. Using this information in (8.27) and combining it with (
8.26),
Z 1" Z x #
1−x x−y
ζ( f, α) ≤ + dy dx
0 2 0 2
Z 1 "Z 1 " # Z 1 #
1−y y − x
= dy + dy dx
0 0 2 x 2
(8.28) = ζ(u, α),
where u stands for the uniform root taking constant value 1/2 on [0, 2]. Simple integration
reveals that ζ(u, α) = 1/3.
Q.E.D.
Lemma 8.9. Given that P( f ) is of the form (8.20), Axiom 1 is satisfied if and only if α ≤ 1.
Proof. Simply inspect (8.19).
Q.E.D.
Lemma 8.10. Given that P( f ) is of the form (8.20), Axiom 2 is satisfied if and only if α ≥ 0.25.
Proof. Consider a configuration as given in Axiom 2: a symmetric distribution made out of

three basic densities. By symmetry, the side densities must share the same root; call this f ∗ .
Let p denote their (common) population mass and m their (common) difference from their
means to their lower support. Likewise, denote the root of the middle density by g∗ , by q its
population mass, and by n the difference between mean and lower support. As in the proof
of Lemma 8.2, we may decompose the polarization measure (8.20) into several components.
First, there are the “internal polarizations” of the middle density (Pm ) and of the two side
densities (Ps ). Next, there are various subtotals of effective antagonism felt by members
of one of the basic densities towards another basic density. Let Ams denote this when the
“origin” density is the middle and the “destination” density one of the sides. Likewise, Asm
is obtained by permuting origin and destination densities. Finally, denote by Ass the total
effective antagonism felt by inhabitants of one side towards the other side. Observe that
each of these last four terms appear twice, so that (writing everything as a function of λ),
overall polarization is given by
(8.29) P(λ) = Pm + 2Ps (λ) + 2Ams (λ) + 2Asm (λ) + 2Ass (λ).
Compute these terms. For brevity, define for any root h,
Z 1 (Z 1 Z 1 )
ψ1 (h, α) ≡ h(x)1+α
h(y)(1 − y)dy + h(y)(y − x)dy dx
0 0 x
and
Z 1
ψ2 (h, α) ≡ h(x)1+α dx.
0
Now, using Lemmas 8.6 and 8.7, we see that
Ps (λ) = 4kp2+α (mλ)1−α ψ1 ( f ∗ , α),
while
Ams (λ) = 2kdq1+α pn−α ψ2 (g∗ , α).
Moreover,
Asm (λ) = 2kdp1+α q(mλ)−α ψ2 ( f ∗ , α),
and
Ass (λ) = 4kdp2+α (mλ)−α ψ2 ( f ∗ , α),
(where it should be remembered that the distance between the means of the two side densities
is 2d). Observe from these calculations that Ams (λ) is entirely insensitive to λ. Consequently,
feeding all the computed terms into (8.29), we may conclude that
" #
d −α q
P(λ) = C 2λ 1−α
+ ψ( f , α)λ { + 2} + D,
∗
m p
where C and D are positive constants independent of λ, and
ψ2 ( f ∗ , α)
ψ( f ∗ , α) =
ψ1 ( f ∗ , α)
by construction; see (8.24) in the statement of Lemma 8.8. It follows from this expression that
for Axiom 2 to hold, it is necessary and sufficient that for every three-density configuration
of the sort described in that axiom,
" #
d −α q
(8.30) 2λ 1−α
+ ψ( f , α)λ
∗
+2
m p
must be nonincreasing in λ over (0, 1]. An examination of the expression in (8.30) quickly
shows that a situation in which q is arbitrarily close to zero (relative to p) is a necessary
and sufficient test case. By the same logic, one should make d/m as small as possible. The
disjoint-support hypothesis of Axiom 2 tells us that this lowest value is 1. So it will be
necessary and sufficient to show that for every root f ∗ ,
(8.31) λ1−α + ψ( f ∗ , α)λ−α
is nonincreasing in λ over (0, 1]. For any f ∗ , it is easy enough to compute the necessary and
sufficient bounds on α. Simple differentiation reveals that
(1 − α)λ−α − αψ( f ∗ , α)λ−(1+α)
must be nonnegative for every λ ∈ (0, 1]; the necessary and sufficient condition for this is
1
(8.32) α≥ .
1 + ψ( f ∗ , α)
Therefore, to find the necessary and sufficient bound on α (uniform over all roots), we need
to minimize ψ( f ∗ , α) by choice of f ∗ , subject to the condition that f ∗ be a root. By Lemma 8.8
, this minimum value is 3. Using this information in (8.32), we are done.
Lemma 8.11. Given that P( f ) is of the form (8.20), Axiom 3 is satisfied.
Proof. Consider a symmetric distribution composed of four basic densities, as in the

statement of Axiom 3. Number the densities 1, 2, 3 and 4, in the same order displayed
in Figure 8.2. Let x denote the amount of the slide (experienced by the inner densities) in the
axiom. For each such x, let d jk (x) denote the (absolute) difference between the means of basic
densities j and k. As we have done several times before, we may decompose the polarization
of this configuration into several components. First, there is the “internal polarization” of
each rectangle j; call it P j , j = 1, 2, 3, 4. [These will stay unchanged with x.] Next, there
is the total effective antagonism felt by inhabitants of each basic density towards another;
call this A jk (x), where j is the “origin” density and k is the “destination” density. Thus total
polarization P(x), again written explicitly as a function of x, is given by
4
X XX
P(x) = Pj + A jk (x)
j=1 j k6= j
so that, using symmetry,

(8.33) P(x) − P(0) = 2{[A12 (x) + A13 (x)] − [A12 (0) + A13 (0)]} + [A23 (x) − A23 (0)]
Now Lemma 8.7 tells us that for all i and j,
Ai j (x) = ki j di j (x),
where ki j is a positive constant which is independent of distances across the two basic
densities, and in particular is independent of x. Using this information in (8.33), it is trivial
to see that
P(x) − P(0) = A23 (x) − A23 (0) = ki j x > 0,
so that Axiom 3 is satisfied.
Given (8.20), Axiom 4 is trivial to verify. Therefore Lemmas 8.9, 8.10 and 8.11 complete the
proof of the theorem.
Q.E.D.
8.6 Polarization and Inequality
Our measurement of polarization is given by
Z Z
Pol = f (x)1+α f (y)|y − x|dydx,
where α lies between 0.25 and 1.

Compare with the Gini coefficient / fractionalization index:
Z Z
Gini = f (x) f (y)|y − x|dydx,
It’s the value of α that makes all the difference. The lower bound on α comes from Axiom 2,
which directly runs against inequality. Otherwise the other axioms are broadly in agreement
in inequality.
Note by the way that there also must be an upper bound on α. For if this were not the case,
identification effects will swamp alienation effects and global squeezes in the distribution
will increase polarization, which makes no sense. This is the job of Axiom 1.
8.7 Three New Properties
1. Bimodality. Polarization maximal for bimodal distributions, but defined of course over all
distributions.
2. Globality. The local merging of two “groups” raises polarization if there is a third “group”
of significant size, but lowers conflict in the absence of such a group.
3. Nonlinearity. Same direction of population or income movements may cause polarization
to go down or up, depending on context.
Density
Income
3. Nonlinearity. Same direction of population or income movements may cause polarization

Density
Income
3. Nonlinearity. Same direction of population or income movement may cause polarization

Density
Income
8.8 Other Remarks on the Measure
8.8.1 More on α. As we’ve discussed, there is a family of possible values of α, subject to the
lower and upper bounds of 0.25 and 1. We’ve already explained where these bounds come
from. The bounds can be narrowed further using additional axioms, though it is unclear
how compelling the additional steps are. Here is an example of the axiomatic approach:
Axiom 5. If p > q but p − q is small and so is r, a small shift of mass from r to q cannot reduce
polarization.
Diagrammatically, go from here:

Density
r p q
Income
to here:
Density
Income
Theorem 8.2. Under the additional Axiom 5, it must be that α = 1, so the unique polarization
measure that satisfies the five axioms is proportional to
Z Z
f (x)2 f (y)|y − x|dydx.
In a later section, we argue that there are strong behavioral reasons to focus on α = 2.
8.8.2 Scaling. Note that our main theorem states that polarization must be proportional to
a particular family of measures. There is therefore a scale factor here which can be employed
as we please. While this is not entirely precise, we can exploit this degree of freedom to
make the measure scale-free. Normalize the measure by µα−1 , where µ is mean income.
This procedure is equivalent to one in which all incomes are normalized by their mean.
Esteban and Ray (1994) begin, in contrast, by using the log of incomes and imposing axioms
on this variable.
8.8.3 Importance of the IA Structure.. Both the axioms and the IA structure needed to
pin down P.
It can be checked that several other candidates satisfy Axioms 1–4. These details are to be
included in future editions of the notes.
8.8.4 Partial Ordering. α varies between bounds, but can vary. So theorem gives us a
partial ordering. Ordinal description of this ordering is an open question.
8.8.5 Identification Windows.. “Identification” here is based on the point density. More
generally, individuals may possess a “window of identification”. Individuals within this
window would be considered “similar” — possibly with weights decreasing with the
distance — and would contribute to a sense of group identity.
At the same time, individuals would feel alienated only from those outside the window.
Thus, broadening one’s window of identification has two effects.
Can capture these two effects somewhat in our seemingly narrower model.
Suppose that each individual at x “perceives” an individual with income y to be at the point
(1 − t)x + ty. Thus the parameter t is inversely proportional to “breadth of identification”.
The “perceived density” of y from the vantage point of an individual located at x is then
!
1 y − (1 − t)x
f
t t
It is easy to see that the polarization measure resulting from this extended notion of
identification is proportional to our measure by the factor t1−α .
It is also possible to directly base identification on the average density over a non-degenerate
window of width w. When we take w to zero, not only is P attained in the limit, but P is a
first-order approximation to Pw , in the sense that ∂Pw /∂w|w→0 = 0.
Nevertheless, the question of identification windows deserves more attention.
8.8.6 Comparing Distributions. Our characterization shows that a comparison across

distributions should depend on average alienation, average identification and on their joint
co-movement. We formalize this very quickly.
R
Let a(y) = |y − x|dF(x), which is the alienation felt by y, and define average alienation ā by
Z Z Z
ā = a(y)dF(y) = |y − x|dF(x)dF(y).
Similarly, f (y)α is the identification felt by y, so define the average α-identification ῑ by

Z Z
α
ῑα ≡ f (y) dF(y) = f (y)1+α dy.
Finally, let ρ be the normalized covariance between identification and alienation:

ρ ≡ covια ,a /ῑα ā.
Then it is easy to verify that

Pα ( f ) = āῑα 1 + ρ .

8.9 Social Polarization
Our polarization measure is easily applicable to ethnolinguistic or religious groupings.

Suppose that there are M “social groups”, based on region, kin, ethnicity, religion... Let
n j be the number of individuals in group j, with overall population normalized to one.
Let F j describe the distribution of income in group j (with f j the accompanying density),
unnormalized by group population. One may now entertain a variety of “social polarization
measures”.
8.9.1 Pure Social Polarization. Consider, first, the case of “pure social polarization”, in
which income plays no role. Assume that each person is “fully” identified with every other
member of his group. Likewise, the alienation function takes on values that are specific to
group pairs and have no reference to income. For each pair of groups j and k denote this
value by ∆ jk . Then a natural transplant of (8.2) yields the measure
M X
X M
(8.34) Ps (F) = j nk ∆ jk .
n1+α
j=1 k=1
Even this sort of specification may be too general in some interesting instances in which
individuals are interested only in the dichotomous perception Us/They. In particular, in these
instances, individuals are not interested in differentiating between the different opposing
groups. Perhaps the simplest instance of this is a pure contest (Esteban and Ray [1999]),
which yields the variant4
M
X
(8.35) P̃s (F) = n1+α
j (1 − n j ).
j=1
4See Reynal-Querol [2002] for a similar analysis. D’Ambrosio and Wolff [2001] also consider a measure of this
type but with income distances across groups explicitly considered.
1
0.9
0.8
0.7
Ethnic fractionalization index
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Ethnic polarization index
If we specialize to the case of α = 1, then a special index of “pure” social polarization is

given by
Figure 8: Ethnic fractionalization versus polarization. Source: WCE.
M X
X M M
X
Pol = n2j nk = n2j (1 − n j ).
j=1 k=1 j=1
This is the measure used by Montalvo and Reynal-Querol, American Economic Review (2005),
a paper that
0.9
we discuss in more detail below. It is instructive to recall the ELF, which is
given by
0.8 M
X
ELF = n j (1 − n j ),
0.7
j=1
and compare
0.6 the two.
Religious fractionalization index
For instance,
0.5
if all groups are of equal size, then polarization peaks when the number of
groups equals 2, and steadily declines thereafter. Fractionalization rises throughout.
0.4
As Montalvo and Reynal-Querol show, it matters empirically too . . .
Guatemala 0.3 and Sierra Leone are examples of countries in which ethnic polarization is high
but ethnic fractionalization is low. Nigeria and Bosnia are examples of countries in which
religious polarization
0.2 is high but religious fractionalization is low.
Once the 0.1two extremes — pure income polarization and pure social polarization — are
identified, we may easily consider several hybrids. As examples, consider the case in
which notions
0 of identification are mediated not just by group membership but by income
similarities 0as well, while
0.2 0.4
the antagonism equation0.6 0.8
remains
Religious polarization index
1
untouched. Then we 1.2get what
Figure 9: Religious fractionalization versus polarization. Source: ET.

one might call social polarization with income-mediated identification:

M
X Z
(8.36) Ps (F) = (1 − n j ) f j (x)α dF j (x).
j=1 x
One could expand (or contract) the importance of income further, while still staying away
from the extremes. For instance, suppose that — in addition to the income-mediation of
group identity — alienation is also income-mediated (for alienation, two individuals must
belong to different groups and have different incomes). Now groups have only a demarcating
role — they are necessary (but not sufficient) for identity, and they are necessary (but not
sufficient) for alienation. The resulting measure would look like this:
M XZ Z
X
(8.37) P (F) =
∗
f j (x)α |x − y|dF j (x)dFk (y).
j=1 k,j x y
Note that we do not intend to suggest that other special cases or hybrids are not possible, or
that they are less important. The discussion here is only to show that social and economic
considerations can be profitably combined in the measurement of polarization. Indeed,
it is conceivable that such measures will perform better than the more commonly used
fragmentation measures in the analysis of social conflict. But a full exploration of this last
theme must await future research (though see the chapter on ethnic conflict below).
8.10 Empirics Revisited: Polarization, not Fragmentation?
We now discuss the recent results of Montalvo and Reynal-Querol (2005). They employ the
same basic specification as Fearon-Laitin (2003) and others, but this time with polarization
instead of fractionalization indices.
They study 138 countries over 1960–1995. The dependent variable is incidence of a civil
war over a five year period. They use what is known as the PRIO25 criterion for civil war,
at least 25 yearly deaths. (Refine this in later versions.) Their explanatory variables include
per-capita income, population size, terrain (proxy for ease of insurgency), primary exports
(proxy for payoff in event of victory), democracy indicators, and of course indices of ethnic
or religious polarization
First run a logit of civil war on ethnic fractionalization. Table 1 reports on the results.
Observe how fractionalization matters in the first column but loses significance completely
as variables such as per-capita income are included. This is the standard result that we had
described earlier.
Now for the logit using ethnic polarization. Table 2 reports. Ethnic polarization is not just
significant through all the variants; the effect is pretty big too. For instance, if polarization is
raised from 0.51 (the average) to 0.95 (Nigeria) the predicted probability of conflict doubles.
[An increase by one standard deviation (0.24) raises conflict probability by 50%.]
Now try the same logit with religious variables instead. Table 3 shows that just like
ethnic fractionalization, religious fractionalization starts out significant but matters quickly
degenerate when additional controls are thrown in. This finding is to be contrasted with
[1] [2] [3] [4]

EthFrac 0.81 0.22 -0.18 0.49
(2.04) (0.53) (0.16) (0.97)
LogPcGdp -0.62 -0.76 -0.79 -0.93
(5.07) (5.90) (5.96) (5.40)
Constant 2.47 -0.42 -0.18 1.57
(2.47) (0.38) (0.16) (0.94)
LogPop 0.46 0.46 0.35
(6.75) (6.03) (3.69)
PrimExp 0.25 0.50
(0.26) (0.48)
Mountains 0.00
(1.67)
NonContiguous -0.20
(0.61)
Democracy 0.49
(1.87)
Pseu R2 0.07 0.15 0.15 0.14

Obs 860 860 840 741
Table 1. Ethnic Fractionalization and Conflict
what happens when a religious polarization variable is employed (Table 4). Religious
polarization stays significant through the various specifications, just as ethnic polarization
did.
These observations are robust to several different specifications. Ethnic polarization is
significant when entered into same regression with ethnic fractionalization; the latter is not.
The same is true if a measure of ethnic dominance (Collier (2001) and Collier and Hoeffler
(2002)) is used instead. Both these observations are still true if “ethnic” is replaced by
“religious”.
The analysis is also robust to the use of different datasets. The World Christian Encyclopedia
is used here to construct ethnic polarization indices. Alternatively, the Encyclopedia
Britannica or the Atlas Nadorov Mira could be used. My main worry here is that I am not
sure how the authors have constructed their ethnic groupings. Clearly they ahev aggregated
some of the highly disaggregated information. But what if they’ve done so in some way
that’s endogenous?
The results are also robust to “joint indices” of ethnic and religious polarization.[Measure
along each dimension, pick the max.] Finally, it appears to be robust to alternative definitions
of civil war.
[1] [2] [3] [4]

EthPol 1.56 1.95 1.98 1.82
(3.31) (3.76) (3.71) (3.23)
LogPcGdp -0.71 -0.77 -0.78 -0.93
(6.16) (6.53) (6.57) (5.50)
Constant 2.65 -1.56 -1.43 -0.93
(3.01) (1.47) (1.27) (0.16)
LogPop 0.49 0.48 0.38
(7.15) (6.46) (4.33)
PrimExp -0.09 0.17
(0.09) (0.16)
Mountains 0.00
(1.13)
NonContiguous -0.00
(0.00)
Democracy 0.41
(1.58)
Pseu R2 0.09 0.17 0.17 0.16

Obs 860 860 840 741
Table 2. Ethnic Polarization and Conflict
8.11 Next Steps
Let’s return to the old question: is ethnicity primordial or instrumental as a determinant of

conflict?
An economist’s instincts suggests that these things are essentiall;y instrumental but I suppose
one never knows . . .
Recall that the findings on per-capita income certainly support an instrumentalist position
on civil war. That suggests a test for ethnic instrumentality: see if economic differences
across groups predicts conflict. But there is an important and fundamental reason why such
a test may be problematic.
Two kinds of economic conflict: “vertical” versus “horizontal”
The “vertical war” certainly exists but is harder to spot and infrequently delineated by
ethnicity. Caste is a good counterexample. With ethnicity and religion, the conflict is
often horizontal: attacks on competing businesses, reduction of labor supply, reallocation of
specific public goods. Listen to Horowitz again:
“In study after study, it has been assumed that ethnic relations are necessary
relations between superiors and subordinates . . . In fact, many ethnic
groups are enmeshed in a system of subordination. But the relations of
[1] [2] [3] [4]

RelFrac 1.41 0.53 0.35 0.92
(2.31) (0.76) (0.49) (1.17)
LogPcGdp -0.61 -0.84 -0.87 -1.03
(4.91) (5.75) (5.85) (5.27)
Constant 1.53 -1.24 -1.15 0.45
(1.42) (0.97) (0.86) (0.25)
LogPop 0.50 0.51 0.41
(6.41) (5.88) (4.09)
PrimExp 0.63 1.15
(0.61) (1.04)
Mountains 0.01
(2.17)
NonContiguous 0.10
(0.31)
Democracy 0.36
(1.29)
Pseu R2 0.10 0.16 0.16 0.16

Obs 853 853 833 734
Table 3. Religious Fractionalization and Conflict
many other ethnic groups — on a global scale, most ethnic groups — are
not accurately defined as superior-subordinate relations . . .
Unlike ranked groups, which form part of a single society, unranked
groups constitute incipient whole societies. It is not so much the politics of
subordination that concerns them, but rather the politics of inclusion and
exclusion.”
How then to augment our polarization measure for wealth differences across and within
groups? The answer may depend on the observer’s feel for the sort of conflict that is relevant.
If conflict is “vertical”, income differences across groups are conducive to conflict, and so is
income homogeneity within groups
On the other hand, if conflict is “horizontal”, income or occupational similarities across
groups may drive conflict and so might income inequality within groups (the buying of
“conflict labor”, as with Dalits in the Gujarat carnage). I will come back to these matters
later.
The interaction of economics and ethnicity creates new conceptual challenges for the
measurement of polarization.
[1] [2] [3] [4]

RelPol 1.09 0.71 0.65 1.06
(2.93) (1.71) (1.50) (2.20)
LogPcGdp -0.57 -0.76 -0.78 -0.98
(4.46) (5.22) (5.26) (5.08)
Constant 1.17 -1.93 -1.85 0.17
(1.10) (1.52) (1.40) (0.10)
LogPop 0.49 0.50 0.39
(6.36) (5.75) (3.94)
PrimExp 0.41 0.93
(0.39) (0.84)
Mountains 0.01
(2.12)
NonContiguous 0.16
(0.47)
Democracy 0.35
(1.26)
Pseu R2 0.10 0.17 0.17 0.17

Obs 853 853 833 734
Table 4. Religious Polarization and Conflict
8.12 Summary So Far
1. Several authors, notably Samuel Huntington, have argued for cultural explanations of
economic development (or lack thereof).
2. Extending this line of thinking, authors such as Donald Horowitz have suggested links
between conflict and ethnic differences.
3. But statistical studies that employ a well-known measure of ethnic and religious
fragmentation show no links with conflict. [Though there are links with economic growth.]
4. In this chapter, I argue for the use of a measure very different from fragmentation — a
polarization index.
5. The measure has a philosophical foundation — the identity-alienation framework — which
may turn out to be useful in other applications.
6. I then discuss an empirical study which uses this polarization measure to exhibit a robust
and positive relationship between (ethnic or religious) polarization and the incidence of
conflict.
7. This does not mean that we buy Huntington on the primordial nature of cultural
differences. An economic war may still be waged behind the ethnic veil.
8.13 A Behavioral Approach to Polarization
No measurement theory can do justice to the claim that there is indeed a link between
polarization and conflict.
Two potential links:
(1) Empirical: fit the measures to the data and see they are related to measured indicators
of social conflict.
(2) Theoretical: write down a “natural” theory of conflict in which the level of
equilibrium conflict is related to distributional polarization.
Here, we concentrate on the latter (with an eye on the former, of course).
8.14 Benchmark Model of Conflict
A unit measure of individuals, situated in G groups. ni : number of individuals in group i,

so that G
i=1 ni = 1.
P
ui j : utility derived by a member of group i if issue j is chosen by society. uii > ui j for all i, j
with i , j.
ri : resources expended per-capita by group i, so total is ni ri .

[Will take a closer look at this later.]
PG
Let R ≡ i=1 ni ri . R is our measure of societal conflict.
Per-capita cost of supplying effort r is c(r): continuous, increasing, smooth, strictly convex,
with c0 (0) = 0.
8.15 Equilibrium
p j : the probability that issue j will be chosen. A member of group i who expends ri gets
expected payoff given by
G
X
(8.38) p j ui j − c(ri )
j=1
To complete the description of the basic model, we assume that

n jr j n jr j
(8.39) p j = s j ≡ PG =
n k rk R
k=1
for all j = 1, . . . , G, if R > 0 (any probability vector otherwise).
(r∗1 , . . . , r∗G ) is an equilibrium if for every i the maximization of the expression in (8.38), subject
to (8.39), is well-defined, and ri solves this problem.
Equilibrium conflict is R∗ = G ∗
P
i=1 ri ni .
The equilibrium resource shares are s∗i = r∗i ni /R∗ .
Neglect internal free-rider problem.
Also assume that groups cannot commit to support issues that are not their favorite.
8.16 Special Cases
(1) Pure Contests: ui j = 0 for all i, j, with i , j, and uii = 1 for all i.
(2) The Line: There is an ordering 1, 2, . . . , G of the groups such that for all i and j with i < j,
ui j ≤ ui,j+1 (if j < G) and ui j ≥ ui+1, j .
ace*0.2in
ui+1,j
uij
ui,j+1
8.17 Background Results
1. Best Response Condition. For each group i, let vi j ≡ uii − ui j for all j.
If r j > 0 for some j , i, then ri > 0, and is given by
G
X
si s j vi j = c0 (ri )ri
j=1
2. Existence and Uniqueness. An equilibrium exists. If, in addition, c000 (r) ≥ 0 for all r, then
the equilibrium is unique.
Remark. c000 (r) ≥ 0 is “necessary” for uniqueness.
8.18 Special Case: Quadratic Costs
Let c(r) = 21 r2 for r ≥ 0.

First order conditions from lemma:
G
X
si s j vi j = c0 (ri )ri
j=1
Rewrite for the quadratic case:

G
X c0 (ri )R ri R si R2
s j vi j = = = 2 .
ni ni ni
j=1
Cross-multiply:
G
X
s j n2i vi j = si R2 .
j=1
G
X
s j n2i vi j = si R2 .
j=1
Form the matrix

n21 v11 n21 v12 ... n21 v1G
 
.. .. .. ..
 
 

 . . . . 

n2i vi1 n2i vi2 ... n2i viG
 
 
.. .. .. ..
 
. . . .
 
 

n2G vG1 n2G vG2 . . . n2G vGG

Then R (squared) is the unique real eigenvalue of this matrix.

[The associated share vector is the unique positive eigenvector (on the simplex) correspond-
ing to this eigenvalue.]
Remark. Observe the squaring of the population coefficients. This is related to polarization
in a way that we will explain later.
8.19 Connections with Polarization
Several connections between behavioral model and the axiomatic approach:
1. Bimodality. Both conflict and measured polarization are maximal for symmetric bimodal
distributions.
2. Globality. The local clustering of two groups raises conflict if there is a third group of
significant size, but lowers conflict in the absence of a third group. Same true of polarization.
3. Nonmonotonicity. Start with a uniform distribution of population across G groups, where

G ≥ 4. Transfer population mass from one of the groups to the others, until a uniform
distribution over G − 1 groups is obtained. Then conflict — and polarization — are higher
at the “end” of this process, but may go down in the “intermediate” stages.
8.20 IsoElastic Costs: c(r) = (1/β)rβ
Use isoelasticity to rewrite first-order conditions:
β
X
si s j vi j = ri
j
Manipulating, we obtain
β
ni
X
s2i s j vi j = si Rβ .
si
j
and adding over all groups i,
P P n i β
Rβ = i j si s2i s j vi j
Compare this result:

P P n i β
Rβ = i j si s2i s j vi j
with axiomatically derived polarization measure for the discrete case:

P = i j n1+α
P P
i
n j vi j
Pretty close for the case α = 1.
Main difference: former is not a closed-form solution, because there are endogenous variables
in it. This is the additional richness imparted by a behavioral model.
8.21 Four Research Questions
A. Group Formation and Interaction.
In many situations, groups are effectively given — men and women, Hindus and Muslims,
importers and exporters ...
In others, there is a wider dispersion of preferences and (sharp) groups arise as a result of
explicit membership decisions — political parties, trade unions, environmental coalitions.
Can be modeled as a two-stage game. For theory, see Bloch [1996, 1997], Ray and Vohra
[1997, 1999, 2001], Konishi and Ray [2003], Yi [1997] ...
[1] In “stage 1”, individuals, possibly endowed with widely varying characteristics, form
groups. (They take “similar” actions, form coalitions or clubs, write binding agreements...)
[2] In “stage 2”, groups “interact” in the way described earlier.
For some recent attempts for “conflict games” see Bloch and Sanchez (2003), Esteban and
Sákovicz (2003), Tan and Wang (2000).
B. Group Salience.
Return to the case of given groups. Many intersecting dimensions: what determines which
group classification is salient?
To some extent, the answer must depend on the set of available policies.
One might say that a society is polarized if the average resistance over a set of policies is high.
Can apply this notion quite easily once we fix a space of policies and a measure over that
space.
E.g., look at a very simple policy x, for which the winners are to one side of the income level
x and the losers are on the other.
Resistance to policy = 1 - F(x)
Winners Losers
R equate polarization to average resistance over a distribution G of policies, we get:

If we
P = min{F(x), 1 − F(x)}dG(x).
Might argue that G is uniform or even equal to F. Still begs the question: what determines
the space of policies?
C. Effort and Money.
Behavioral model looks pleasantly general, but in fact the explicit introduction of income
(or wealth) poses new challenges.
How are resources expended in the lobbying process?

[In a “perfect” democracy where all questions are solved by referenda, this problem becomes
irrelevant.]
The poor use effort; the rich use money.

The political system determines a political exchange rate between effort and money.
Individual characteristics (abilities, access, wealth) determine an economic exchange rate

between effort and money.
The cutoff for different types of participation (effort vs. money) will depend on the political
system.
The structural features of such a model will impose useful restrictions on empirical analysis.
Typically, instances of open conflict will be observable while the use of money to influence
policies will not be.
Suggests the use of modified polarization measures in which the symmetry between different
groups is broken if there are wealth differences.
D. Lobbies as Signals.
Based on Esteban and Ray (2006).

[1] Governments play a role in the allocation of resources.
[2] Governments lack information — just as private agents do — regarding which sectors
are worth pushing in the interests of economic efficiency.
[3] Agents (sectoral interests, industrial confederations, R&D coalitions ...) lobby the
government for preferential treatment.
[4] A government — even if it honestly seeks to maximize economic efficiency — may be

confounded by the possibility that both high wealth and true economic desirability create
loud lobbies.
Connects inequality and lobbying to resource allocation.
8.22 Summary
I have discussed two approaches to study of polarization and conflict.
The axiomatic approach delivers a new set of measures which may be useful in empirical
work.
The behavioral approach complements the axiomatic approach by explicitly laying down a
model of conflict.
While the axiomatic approach is possibly of greater empirical relevance, the behavioral
approach cannot be dispensed with as a conceptual check on the axioms.
In particular, issues of group formation, group salience, the use of alternative forms of
resources in a conflictual process, and role of lobbying as a signaling device can be usefully
analyzed under the behavioral approach.
8.23 Other Issues in the Theory of Conflict
8.23.1 Public Versus Private Goods. So far we have assumed that conflict takes place
over the allocation of public goods. One might also be interested in situations in which the
good to be allocated is private or partly private. For simplicity, suppose that there are only
two groups, with population sizes n1 and n2 summing to one. We also suppose that the cost
function is isoelastic, c(r) = (1/α)rα for α ≥ 2 (thus satisfying our third-derivative condition).
There is a budget of G which is up for grabs. Assume that a fraction λ of this is public, while
the remaining fraction is private and divided equally among the winning group. Group i
then chooses ri to maximize
ni ri G

λG + (1 − λ) − c(r).
R ni
h i
Define P(n) = λG + (1 − λ) nGi ; then the FOC for this problem is
 ni n2i ri 
 
P(ni )  − 2  = rα−1 i ,
R R
or
rα−1
2 i
(8.40) P(ni )ni n j = R .
rj
Similarly,
rα−1
j
(8.41) P(n j )ni n j = R2
.
ri
Raising both sides of (8.41) to the power α − 1 and multiplying with (8.40), we see that
α(α−2)
(8.42) P(ni )ni n j [P(n j )ni n j ]α−1 = R2α r j .
If α = 2, we see right away that
(8.43) R4 = P(ni )n2i P(n j )n2j ,
and it is very easy to now see that this expression is maximized at the symmetric bipolar
distribution in which ni = n j = 1/2. Thus in the quadratic case the degree to which the
prize is public or private matters not at all in determining which distribution is the most
conflictual.
When α , 2 an extended version of (8.43) holds, which we now derive. Use equation (8.42)
and the assumption that α , 2 to write
i1/α(α−2)
r j = P(ni )P(n j )α−1
h
(8.44) (ni n j )1/(α−2) R−2/(α−2) .
Of course, a parallel equation holds for ri :

i1/α(α−2)
ri = P(ni )α−1 P(n j )
h
(8.45) (ni n j )1/(α−2) R−2/(α−2) .
Multiplying both sides of (8.44) by n j and both sides of (8.45) by ni , and adding the two
equations, we see that
h i(α−1)/α(α−2) n o
R = R−2/(α−2) P(ni )P(n j ) (ni n j )1/(α−2) P(ni )−1/α n j + P(n j )−1/α ni ,
and after manipulating (quite) a bit, we conclude that

i(α−1)/α oα−2
Rα = P(ni )P(n j )
h n
(ni n j ) P(ni )−1/α n j + P(n j )−1/α ni
h i1/α n oα−2
(8.46) = P(ni )ni P(n j )n j (ni n j )(α−1)/α P(n j )1/α n j + P(ni )1/α ni
and this is the appropriate extension of (8.43), which was derived for α = 2 (in fact, just put
α = 2 in (8.46) to recover (8.43)).
It can be shown that for α > 2, this function is maximized at ni = n j = 1/2.5 Details relegated
to Digression below.
Digression. Do you really want to read this? Well, carry on if you must. We are going to show that
the last line in (8.46) is maxed at ni = n j = 1/2. There are three terms in this last line, each raised to
different powers. It is very easy to verify that the first and second of these, P(ni )ni P(n j )n j and ni n j ,
are each maximized when ni = n j = 1/2. The third term (which is raised to the power α − 2) takes a
bit more work. It will suffice to show that the expression
i1/α
P(n)1/α n = λGnα + (1 − λ)Gnα−1
h
is concave in n. It is actually trivial to verify this either for λ = 0 or λ = 1 but the intermediate
steps require some calculus. Differentiate the expression above with respect to n to obtain the first
derivative proportional to
1h α i(1−α)/α h
λn + (1 − λ)nα−1 λαnα−1 + (1 − λ)(α − 1)nα−2
i
∆(n) =
α
i(1−α)/α (1 − λ)(α − 1) (α2 )α (1−α)/α
λ n−α λnα + (1 − λ)nα−1 n 1−α λnα + (1 − λ)nα−1
h
= +
α
i(1−α)/α (1 − λ)(α − 1) h i(1−α)/α
λnα/(α−1) + (1 − λ)n1/(α−1)
h
= λ λ + (1 − λ)n−1 +
α
(1 − λ)(α − 1) −1
" #
h i(1−α)/α
= λ+ n λ + (1 − λ)n−1 .
α
5I don’t believe this result holds when α < 2.

To complete the argument we must show that the very last line above is negatively related to n. To
do this it suffices to show that the derivative of
(1 − λ)(α − 1)
" #
λ+ x [λ + (1 − λ)x](1−α)/α
α
with respect to x is nonnegative. That derivative is given by
(1−α)/α (1 − λ)(α − 1) (1 − λ)(α − 1) 1 − α
" #
[λ + (1 − λ)x] + λ+ x [λ + (1 − λ)x](1−2α)/α (1 − λ)
α α α
(1−λ)(α−1) 
 
(1 − λ)(α − 1)  λ + α x
= [λ + (1 − λ)x](1−α)/α 
 
1−
 
α λ + (1 − λ)x 
 
 
≥ 0.
We’ve therefore shown that the bipolarization result is quite robust when it comes to
explaining the intensity of conflict, conditional on it happening in the first place. Whether
the good to be fought over is public or private really does not matter too much.
8.23.2 Conflict Initiation. The italicized phrase in the last paragraph is important, how-
ever. Highly polarized situations may not exhibit any conflict to begin with, because
everyone knows that the conflict is going to be very costly if it happens. Therefore, a little
bit more attention has to be paid to theories that attempt to relate conflict to polarization.
Let us explore these matters a bit further as they have interesting implications for the identity
of the initiating party. Suppose again that there are two groups, and that group i receives
a peacetime payoff in which it obtains a share si of the good. Then it his easy to see thati its
peacetime payoff is simply given by si P(ni ), where we define P(n) = λG + (1 − λ) n as in
G
the previous section. It follows that group i will want to engage in conflict provided that
ni ri
(8.47) P(ni ) − c(ri ) > si P(ni ),
R
where ri and R are to be interpreted as equilibrium values in the ensuing conflict.
Once again adopt the isoelastic specification and recall the first-order condition (8.40) with
respect to ri . But write it in a form that we used for our conflict and distribution study:
(8.48) P(ni )pi p j = rαi ,
where pi (and likewise p j ) is just the win probability or conflict share ni ri /R. We may use
this to obtain a quasi-closed-form for the net payoff from conflict:
1

(8.49) P(n)pi − c(ri ) = P(ni ) pi − pi (1 − pi ) .
α
Now we proceed to an analysis of conflict initiation. Consider the equivalent formulation
of the first-order condition for group i (which essentially reproduces (8.40)):
rα−1
i
(8.50) P(ni )ni n j = R2 .
rj
If the good in question is purely public, then we may simply normalize P(n) = 1, so that
rα−1
i
ni n j = R2 .
rj
Of course, exactly the same condition applies to r j , and it therefore follows that ri = r j = R,
and so pi = ni . Using this information in (8.49) above and invoking (8.47), we obtain the
following condition for conflict initiation:
1
ni (1 − ni ) > si .
ni −
α
It is easy to see that the left-hand side of this expression is strictly increasing in ni ,6, and that
it must therefore cross the right-hand side from below (provided si lies strictly between 0
and 1). We conclude then that large groups initiate conflict when the battle is over public
goods.
Place si at 1/2 for symmetry. Observe that the condition never holds when ni = 1/2. This
is the most polarized case but it does not exhibit conflict. On the other hand, once conflict
breaks out it is monotonically related to polarization, as discussed in the previous section.
So equilibrium conflict jumps up at the threshold and then declines again monotonically to
zero as ni → 1.
√
An interesting point to note is that ni > s is always sufficient for conflict (simply make
the LHS as small as possible in α by putting α = 1). This is a more general point that
is independent of the cost function.7 In any case the important point is that large groups
initiate in the public goods case.
This observation is, however, reversed when the good at stake is private. To see this, recall
the first-order condition for group i, (8.50), and normalize P(n) = 1/n to see that
rα−1
i
n j = R2 .
rj
Dividing by the corresponding first-order condition for group j, we may conclude that
n 1/α
ri j
= ,
rj ni
which captures part of the Olson intuition that small groups lobby more per-capita.8
Therefore the probability that i wins the conflict is given by
(α−1)/α
ni ri n
(8.51) pi ≡ = (α−1)/αi ,
n i ri + n j r j n + n
(α−1)/α
i j
6The reason is that the derivative of n(1 − n) never exceeds 1.

7To prove this, note that the more general first-order condition is just n n = R2 c0 (r )/r for both i and j, so it is
i j i j
still true that ri = r j = R. It follows that c0 (R)R = ni n j . Now recall the conflict initiation condition (8.47), and note
that by convexity c0 (R)R ≤ c(R). Therefore a sufficient condition for (8.47) to hold is that ni − ni n j > si , which
proves the claim.
8This is, however, far from a full examination of the Olson paradox as we also have to allow for free-riding
within the group. For a model that does this, see Esteban and Ray (APSR 2001).
so in particular large groups still have a higher win probability.

Finally, invoke (8.49) and (8.47) to write down the condition for conflict initiation in the
private goods case, which is
1

pi − pi (1 − pi ) > si ,
α
where pi is now given by (8.51).
In contrast to si = 1/2, the symmetric case for peacetime is now si = ni . Let us use this
benchmark to understand conflict initiation. Rewriting the condition above with si = ni , we
obtain
1 2 α−1
(8.52) p + pi > ni .
α i α
Combine this inequality and (8.51) to now establish the following observation. There is a
unique population threshold below which a group will initiate conflict.
To prove this claim, recall (8.51) and take a closer look at p (dropping subscripts):
nk
p(n) = ,
nk + (1 − n)k
where k ≡ (α − 1)/α. The function p has an interesting “reverse-logistic” shape. It starts
above the 450 line and at the point n = 1/2 crosses it and dips below. The derivatives at the
two ends are infinite. To check these claims, note that
nk
≥n
nk + (1 − n)k
if and only if n ≤ 1/2 (simply cross-multiply and verify this), and that
knk−1 (1 − n)k−1
p0 (n) = ,
[nk + (1 − n)k ]2
which is infinite both at n = 0 and n = 1. Now recall (8.52) and write it as
(1 − k)p(n)2 + kp(n) > n.
By the arguments just made on derivatives, the LHS starts out higher than the RHS and ends
up lower than the RHS. This means that conflict is preferable for small minorities and not so
for large majorities, in contrast already to the results for public goods.
Indeed, we can strengthen that last argument to show that
(1 − k)p(n)2 + kp(n) < n
for any n ≥ 1/2, that conflict becomes strictly bad for any weak (nonunanimous) majority.
Suppose this is false for some 1 > n ≥ 1/2. By the properties of p already established, we
know that n ≥ 1/2 implies n ≥ p(n), so that
(1 − k)n2 + kn ≥ n,
but this can never happen when n < 1, a contradiction. So conflict can never be preferable
for a weak majority of the population.
It remains finally to show that there is a unique intersection (crossing from above to below)
in the interior. Let n∗ ∈ (0, 1) be an interior solution; then
(1 − k)p(n∗ )2 + kp(n∗ ) = n∗ .
Differentiate to show that the crossing is “from above”.
CHAPTER 9
Inequality and Incentives
In this chapter we study collective action problems (such as team production or the voluntary
provision of public goods) and identify three channels through which inequality affects
incentives and therefore overall efficiency.
The general structure of a team production problem is as follows. There is a group of n
agents, each of whom contributes resources (money, effort) ri . The joint output is given by
(9.1) Y = F(r)
where r = (r1 , . . . , rn ) is a (nonnegative) vector.

Depending on the application, Y could be a public good or a private good.
I now introduce the three channels through which inequality might function:
A. Wealth. Suppose that the team good is public, and that person i has wealth endowment
wi . The person i’s payoff may be written as
Y + u(yi − ri ).
The question then is: how does inequality in wealth affect incentives?
B. Access. Suppose that all agents are identical in endowment, and the good is private. But
it is unequally distributed. Then different individuals have different shares of Y.
In this example, our proxy for inequality is a vector of shares ` ≡ (λ1 , λ2 , λn ), which sum to
unity. For instance, this may be a problem of joint maintenance of an irrigation network, or
of cooperative farming, in which case the shares would proxy the amount of land holdings
of the agent.
How might inequality in access affect incentives?
C. Inequality Tolerance. Attitudes to inequality might affect incentives in situations where
commitment is not possible. To illustrate this problem suppose that there are two regions
A and B which receive funding from an equality-minded central government. Suppose that
region i invests ri and produces yi = f (ri ) (think of yi as regional tax revenues and ri as
resources put into the generation of local revenues).
178 Inequality and Incentives
Suppose that the center must choose transfers ti to each of the regions, satisfying some
budget constraint tA + tB = T. Suppose that it has welfare function
W(yA + tA ) + W(yB + tB ).
If the center can commit to the transfers then in a symmetric optimum it would set ti = T/2
and then each region i would choose ri to
max ti + f (ri ) − ri ;
this would implement the first-best without a problem. However, if the center does not
commit to the transfers then the states move first and choose yi . This creates a dilemma if
the center then picks tA and tB to
max W(yA + tA ) + W(yB + tB ),
with yA and yB given. Ignoring corner solutions, we see that this would entail
yA + tA = yB + tB .
This creates a disincentive problem for the regions. Each region i will anticipate this
compensatory behavior by the center and so choose ti to maximize
yi + ri (yA , yB ).
It is easy to see that this leads to the first-order condition
1 0
f (ri ) = 1,
2
which implies substantial underinvestment relative to the first best.
Clearly, much of this depends on how the center reacts to the choices of yA and yB , which in
turn depends on the degree of egalitarianism felt by the center, which is embodied in W. It
also depends on whether the center plans to “compensate” the regions for their investment
of r. This leads to the general class of questions:
When a principal cannot commit to a reward function, how does her tolerance for inter-agent
inequality affect the incentives of participating agents?
9.1 Inequality in Endowments
I build on a classic paper on the voluntary provision of public goods by Bergstrom, Blume
and Varian (1981). There are n agents, each of whom contributes resources ri to a public
good. Output is given by
(9.2) G = f (r),
where r = i ri , and f is smooth, increasing and concave. (One can take a more general
P
specification but this will do to illustrate our main points.)
The individual utility function is
u(ci ) + G,
where ci = wi − ri for all i, and u is smooth, increasing and strictly concave.
Inequality and Incentives 179
An equilibrium is a vector r∗ such that every individual i choose r∗i optimally, given that
resources contributed by the others is r∗j for j , i.
Proposition 9.1. There exists a unique equilibrium.
Proof. Suppose, on the contrary, that two distinct vectors r and r0 are both equilibria.
Without loss of generality r0 ≥ r. Pick i such that r0i > ri . Because r0i > 0, we must have
(9.3) u0 (wi − r0i ) = f 0 (r0 ),
while
(9.4) u0 (wi − ri ) ≥ f 0 (r).
Because r0i > ri , we have that u0 (wi − r0i ) > u0 (wi − ri ). Using this in (9.3) and (9.4), we must
conclude that
f 0 (r0 ) = u0 (wi − r0i ) > u0 (wi − ri ) ≥ f 0 (r),
so that by the concavity of f , r0 < r. This is a contradiction.
Our next proposition describes an interesting fact about the Bergstrom-Blume-Varian model,
which is that — to the extent that contributions are interior — it is a distribution-neutral
theory.
Proposition 9.2. Suppose that under a wealth distribution w every individual makes strictly positive
contributions. Consider another wealth distribution w0 , with the same aggregate wealth, such that
no individual’s wealth is reduced by more than her contribution under w. Then r = r0 : overall
contributions are unaffected.
Proof. Define i ≡ w0i − wi for every i, and then define r0i ≡ ri + i . Because i i = 0, we have
P
r0 = r. Also, r0i ≥ 0 for all i because ri ≥ −i , by assumption. Now recall that
u0 (wi − ri ) = f 0 (r)
for all i, because ri > 0, so we must conclude that
u0 (w0i − r0i ) = f 0 (r0 )
as well. This means that r0 is an equilibrium under the new wealth distribution, and by
Proposition 9.1, it is the only one.
Proposition 9.3. In any equilibrium, everyone who makes positive contributions has their
consumptions and utilities fully equalized, regardless of their wealth.
Proof. For ri and r j both positive, we must have

u0 (wi − ri ) = u0 (w j − r j ) = f 0 (r),
so that ci = c j and ui = u j .
This is a strange result, and it cannot be broken even with very general utility functions such
as u(c, G). It does break down if f is asymmetric across agents or if the utility function is
heterogeneous across agents.
This motivates a careful study of corner solutions in the voluntary contributions model.
These solutions are not at all exceptional: it is quite reasonable to suppose that some poor
agents do not make contributions, while the richer agents do. So let w be some wealth
distribution. Let I be the set of all agents who make positive contributions: I = {i|ri > 0}.
Observe that small redistributions of wealth within I have no effect on aggregate outcomes,
just as in the neutrality proposition studied earlier.
What about a redistribution from j < I to some i ∈ I?
Proposition 9.4. Greater inequality in wealth either leaves aggregate contributions unchanged, or
pushes them towards higher levels.
Proof. To prove this, it suffices to consider a small transfer from j < I to i ∈ I, such that
the set I remains unchanged. (This follows from Proposition 9.2 on distribution-neutrality.)
I claim that in the new equilibrium, r0 > r. Suppose not; then r0 ≤ r. Now
u0 (wk − r0k ) = f 0 (r0 )
for k ∈ I, k , i. Because f 0 (r0 ) ≥ f 0 (r), we must conclude that r0k ≥ rk for all such k. For i, we
have that
u0 (wi − r0i ) = u(wi + − r0i ) = f 0 (r0 )
P argument, ri − ≥ ri . Putting all this together, we must conclude that

so by 0
P the same
r = k rk ≥ k rk + > r, a contradiction.
0 0
Therefore in the Bergstrom-Blume-Varian model, we must conclude that greater wealth

inequality must lead to greater provision of the public good. Roughly speaking, wealth
inequality raises the contribution costs of noncontributors and lowers the cost to contribu-
tors. The former has no effect: noncontributors are noncontributors and it does not matter
whether their costs go up or not. The latter increases contributions; hence the result.
It should be pointed out that the effect on total (utilitarian) surplus is uncertain. After
all, utilities are concave and disequalizing them through regressive transfers of wealth has
inherently negative consequences. However, sometimes one can argue that overall welfare
goes up. A disequalization of wealth has three effects: (i) it increases rich utilities (positive);
(b) it lowers poor utilities (negative) and (c) it increases public goods supply (positive). To
compute net effects think of a two-person example in which 1 is a contributor and 2 is not.
If we are very close to the contribution threshold for 2, we have that
w1 − r1 ' w2
by the equalization proposition above. Now a disequalization of wealth will have no first-
order effect on the sum of utilities, while it will have a first-order positive effect on public
good provision.
In the next section, we will see an important sense in which Proposition 9.4 may not be
robust.
9.2 Inequality in Access
In this section, we assume that the output is a private good, though with a little bit of clever
renormalization most of the analysis goes through with mixed public-private goods.
The objective is to analyze access inequalities, so we will have the produced output divided in
some given way among the contributing agents. The question is: how does the inequality
in this division affect efficiency?
Our focus will be on an aspect of production that we did not consider in the last section: we
will allow for the inputs of different agents to be complements in producing the joint output.
The maintenance of an irrigation network is a good generic example: the maintenance
of different stretches of an irrigation channel is complementary, and the idea of access
inequalities also makes sense (think of irrigation being important in proportion to the amount
of land being farmed).
Other than complementarities, we keep everything very simple. Individual utility functions
are identical, and taken to be linear both in consumption and in contributions (which, in
keeping with the maintenance example) we will now call effort e. Thus each agent i seeks to
maximize ci − ei , and chooses ei to
(9.5) max λi F(ei , e−i ) − ei
ei
where the notation e−i stands for the vector e with the ith component removed.
An equilibrium is an effort vector e∗ with the property that for every i, e∗i solves (9.5), given
e∗−i .
An effort vector ê is efficient if it maximizes the expression
n
X
Ŝ ≡ F(e) − ei
i=1
over all possible effort vectors. Thus Ŝ is the maximal surplus that can be generated in
the economy. Assume that appropriate end-point conditions hold so that the maximization
problem above is well-defined.
Define the surplus S∗ associated with any equilibrium e∗ by the expression F(e∗ ) − ni=1 e∗i .
P
Obviously, S∗ ≤ Ŝ in any equilibrium. Take Ŝ − S∗ to be our measure of the inefficiency of an

equilibrium.
This definition naturally induces a class of inefficiency measures for a given level of “access
inequality”, as proxied by `. We adopt the following definition: the inefficiency I(`) of access
inequality ` is given by
(9.6) I(`) ≡ min{Ŝ − S∗ | S∗ is the surplus associated with some equilibrium e∗ under `}.
The question that we seek to explore is: which levels of access inequality minimize
inefficiency? Writers such as Mancur Olson have stressed that unequal sharing rules are
good for efficiency, because they minimize the free-rider problem. We may summarize this
intuition in the following
Proposition 9.5. Suppose that F is an increasing, concave function of the sum of efforts: F(e) =
f ( ni=1 ei ) for some increasing differentiable concave f satisfying the Inada endpoint conditions. Let
P
` = (λ1 , . . . , λn ) and `0 = (λ01 , . . . , λ0n ) be two different access inequalities, with the property that
maxi λi > maxi λ0i . Then I(`) < I(`0 ). Moreover, I(`) → 0 as ` converges to any of the unit vectors.
Proof. Let M(`) ≡ maxi λi . Then in any equilibrium e∗ under `, e∗i is positive only if λi = M(`),
and consequently, if E∗ ≡ ni=1 e∗i , E∗ must maximize M(`) f (E) − E with respect to E. Thus I(`)
P
is simply Ŝ − M(`) f (E∗ ) + E∗ . It is easy to check that I(`) is monotonically decreasing in M(`)
and indeed, that I(`) → 0 along any sequence such that M(`) → 1.
Thus under some conditions, inequality of sharing is conducive to efficiency. These

conditions require that output be a function of the sum of efforts. This assumption may
be particularly cogent in the case of lobbying, where the effectiveness of lobbying may be a
function of the sum of monetary contributions.
However, in many production activities, the efforts of different individuals may be
complements in production. The following equally simple proposition considers the extreme
Leontief case, to provide a stark contrast to Proposition 9.5.
Proposition 9.6. Suppose that F is an increasing concave function of the scale of activity, where
scale is determined by equi-proportional contribution of efforts: F(e) = f (mini ei ) for some increasing
differentiable concave f satisfying the Inada endpoint conditions. Let ` = (λ1 , . . . , λn ) and `0 =
(λ01 , . . . , λ0n ) be two different access inequalities, with the property that mini λi > mini λ0i . Then
I(`) < I(`0 ).
Proof. Note that in this case we have a continuum of equilibria for each possible level of
access inequality `. These are characterized as follows.
Define m(`) ≡ mini λi . Then for each `, e∗ is an equilibrium vector of efforts if and only if
e∗1 = e∗2 = . . . = e∗n = e (say), and m(`) f 0 (e) ≥ 1. Thus
I(`) = Ŝ − m(`) f (e) + ne.
It is easy to check that I(`) is a decreasing function of m(`).
Thus in this case, we have exactly the opposite result: equality is conducive to efficiency,
though unlike the case of proposition 9.5, we never obtain full efficiency (it is easy to check
that I(`) is bounded away from zero). We may therefore conjecture that the least inefficient
access inequality varies in some systematic way with the degree of complementarity in
production.
9.2.1 More on Access Inequalities. The analysis above suggests that the degree of
substitution will play a role in determining whether or not inequality is good for efficiency.
To capture this idea, parameterize production so it runs the gamut from perfect to zero
substitutability of effort is to use a CES structure for the vector of efforts. Thus we may take
F to have the form
 n  1−σα
X 
(9.7) F(e) =  e1−σ
i
 ,

i=1
where σ ≥ 0 measures the degree of effort substitution, and α ∈ (0, 1) is a scale parameter.
Direct computation reveals that the maximal surplus Ŝ in this model is given by
α ασ
(9.8) Ŝ = α 1−α (1 − α)n (1−α)(1−σ) .
To compute the Nash equilibrium, note first that for σ > 0, best responses for each i are
strictly positive provided that all other efforts are strictly positive. So let us write down
interior first-order conditions to the maximization problem (9.5), for each i, which we may
rephrase here as follows:
 α
 1−σ
Xn 
(9.9) max λi  e1−σ  − ei
 
ei  j 
j=1
Simple computation reveals the first-order conditions to be

  α+σ−1
Xn  1−σ
σ
(9.10) ei = αλi  1−σ 
e j  ,


 
j=1
and summing over all i in (9.10) and simplifying, we see that

 σ
 1−α
n n
X 1−σ 
 X 1−σ 
e1−σ = α 1−α  λ j σ 

(9.11)

j  
j=1 j=1
We may substitute (9.11) in (9.10) to conclude that

 α+σ−1
 (1−α)(1−σ)
n
1 1 
X 1−σ 
ei = λ σ α 1−α  λ j σ  .

(9.12)

 
j=1
Equipped with these equations, we may now obtain a closed form for the surplus S∗ (`) in
any interior Nash equilibrium under `:
 α+σ−1
 (1−α)(1−σ)  
n n n
α 
X 1−σ  X 1−σ X 1
S∗ (`) = α 1−α  λ j σ  λj σ − α λ jσ  .
 
(9.13)
  

   
j=1 j=1 j=1
The problem that we are interested in thus reduces to the seemingly innocuous maximization
exercise: maximize S∗ (`), as given by (9.13), with respect to `.
It is possible to use the equations (9.8) and (9.13) to run some simple consistency checks.
Propositions 9.5 and 9.6 may be obtained by the taking of appropriate limits in the above
maximization exercise. The more general issue is to calculate the values of ` to intermediate
problems.
We do not have a complete solution to this problem. But here are some observations.
Proposition 9.7. If σ ≥ 1/2, then perfect equality maximizes constrained surplus.
Proof. Let λ = (λ1 , . . . , λn ) be a vector of nonnegative shares summing to unity. We already

have a closed form for the surplus S∗ (λ) in any interior Nash equilibrium, which I write here
slightly differently:
θ
Pn
α j=1 λ j (1
− αλ j )
(9.14) S (λ) = α
∗ 1−α P γ ,
n θ
j=1 λ j
σ , and γ ≡ (1−α)(1−σ) . Notice that γ is only well-defined if σ , 1, which we shall

where θ ≡ 1−σ 1−α−σ
assume in what follows. [The logarithmic case σ = 1 will be discussed at the end of the
proof.]
We will use this expression to establish the proposition. To this end, let us compute the
derivative of S∗ (λ) with respect to any one of the shares, say λi . This derivative is given by
−α
∂α 1−α S∗ (λ) ∆(λi )
= γ ,
∂λi Pn θ
λ
j=1 j
where (after some computation) it can be seen that

θ
Pn
j=1 λ j [1
− γ(1 − αλ j )]
(9.15) ∆(λi ) = θλθ−1
i { Pn θ
} − α(θ + 1)λθi .
λ
j=1 j
Note: we will be done if we can show that whenever λi < λk , then ∆(λi ) > ∆(λk ). For this
would mean that a small transfer from the higher share to the lower share would raise total
surplus. Because the terms within curly brackets in (9.15) are unchanged in this comparison,
it will suffice to show that the function
θ
Pn
j=1 λ j [1 − γ(1 − αλ j )]
θ−1
(9.16) ∆(x) = θx { Pn θ
} − α(θ + 1)xθ
λ
j=1 j
is strictly decreasing in x as x varies over [0, 1]. To this end, differentiate ∆(x) to see that
θ
Pn
j=1 λ j [1− γ(1 − αλ j )]
θ−2
(9.17) ∆ (x) = θ(θ − 1)x
0
{ Pn θ
} − αθ(θ + 1)xθ−1 .
λ
j=1 j
It will suffice to show that this expression is strictly negative for x ∈ (0, 1).
We distinguish between two cases.
Case 1. σ ∈ (1/2, 1). In this case it is easy to see that θ ∈ (0, 1) and that γ < 1. Using this
information in (9.17), it follows right away that ∆0 (x) < 0.
Case 2. σ > 1. In this case it is easy to see that θ < 0 and that γ > 1.1 Using part of this
information (the fact that θ < 0), we see from (9.17) that ∆0 (x) < 0 if
θ
Pn
j=1 λ j [1 − γ(1 − αλ j )]
(θ − 1){ Pn θ
} > α(θ + 1)x,
j=1 λ j
or equivalently, if we can establish the inequality

θ
Pn
j=1 λ j [1 − γ(1 − αλ j )] θ+1 1
(9.18) { } < αx = αx.
θ θ−1
Pn
j=1 λ j
1 − 2σ
However, notice that
θ
Pn
j=1 λ j [1
− γ(1 − αλ j )] α
≤ 1 − γ(1 − α) = − ,
θ σ−1
Pn
j=1 λ j
where use is made of the fact that λ j ≤ 1 for each j. With (9.18), this means that it will suffice
to establish the inequality
α 1
− < αx,
σ − 1 1 − 2σ
or equivalently,
α 1
> αx,
σ − 1 2σ − 1
But this inequality follows from direct inspection, and we are done.
Notice that the logarithmic case σ = 1 remains uncovered. I believe this should be established
by separate computation.
Proposition 9.8. If σ < 1/2, there exist population sizes for which equal division cannot be optimal.
Proof. As a construction for the proof, set up the equal minority problem with m people
receiving equal shares 1/m: call this vector λm . Then it is easy to see that
(9.19) S∗ (λm ) ' mb (m − α),
ασ+σ−1
where b ≡ (1−α)(1−σ) .
Define m∗ by the smallest integer such that

1 − ασ − σ
(9.20) m∗ ≥
1 − 2σ
I will now show that S∗ (λm ) > S∗ (λn ) for any n > m∗ , which proves that perfect equality is
impossible when n exceeds m∗ . To this end, we pretend that m is a continuous variable in
(9.19). It will suffice to show that the derivative of S(m) with respect to m is negative for all
m > m∗ . Differentiating (9.19) with respect to m, we need to show that
(b + 1)m − αb < 0
1To check the observation for γ, be a little careful. First note that when σ > 1, γ can be written as σ+α−1
, and
(σ−1)(1−α)
now examine this.
for all m > m∗ . Using the definition of b and m∗ , this is a matter of simple algebra.
The simulations (see below) show a strange area where two players receive different shares.
It looks subtle. The following proposition takes us part of the way towards resolving these
strange issues.
Proposition 9.9. There can be no more than two distinct positive values of the share in any
constrained optimum. Moreover, there can be no more than one person endowed with the lower of the
two positive values of the share.
Proof. To prove this, recall the expression for ∆(λi ) in (9.15), and note the following:
Lemma 9.1. Let λ be a share vector. If ∆(λi ) and ∆(λk ) are different for two positive values λi and
λk in the share vector, then that share vector cannot be a constrained optimum.
The proof of this lemma is trivial. Simply transfer shares from the one with lower ∆-value
to the one with the higher ∆-value. Since the shares in question are both positive, there is no
constraint to doing this in any “direction” we please.
Given Lemma 9.1, our question boils down to this: for how many distinct and positive
values of x can ∆(x) have the same value? Recall that ∆(x) is defined in (9.16);2 it may be
written as
(9.21) ∆(x) = θxθ−1 − Cxθ ,
where
α(θ + 1) nj=1 λθj
P
C ≡ Pn θ
> 0.
j=1 λ j [1 − γ(1 − αλ j )]
Now we prove the following observation: the function ∆(x), for x ≥ 0 is “single-peaked” (though
not necessarily concave): first rising, then falling.
To show this be a little careful, because the function is not necessarily strictly concave (try
θ > 2). But it is easy to do (details omitted).
Now we are done with the first part of the proposition. Such a function can exhibit the same
value for at most two distinct points in the domain.
To complete the proof, let a denote the smaller of the two positive values of the share. We will
show that if two (or more) persons are given a, we can improve the surplus by transferring
some share from one of them to the other. To this end, think of the share of these two
individuals as variables x and y. “Initially”, x = y = a. Holding all other shares constant, we
may think of the aggregate surplus simply as a function S(x, y). For some > 0, we know
from the mean value theorem for multivariate functions (see, e.g., Hoffman (1975, Section
8.4, Theorem 6)) that
(9.22) S(a + , a − ) − S(a, a) = [S1 (x̂, ŷ) − S2 (x̂, ŷ)]
2It is true that the shares also appear within the curly brackets. But this does not matter as we are simply
comparing for different shares “over the cross-section” and the share vector does not change.
where superscripts denote the appropriate partial derivatives, and x̂ can be chosen to be
strictly larger than ŷ.3 Now we know that
(9.23) S1 (x̂, ŷ) = θx̂θ−1 − C(x̂, ŷ)x̂θ ,
and
(9.24) S2 (x̂, ŷ) = θ ŷθ−1 − C(x̂, ŷ) ŷθ ,
where C(x̂, ŷ) is exactly the same C as before (see definition just following (9.21)), but with
two of the arguments made explicit to remind the reader that x̂ and ŷ enter there as well.
Recall now that
θzθ−1 − C(a, a)zθ
is strictly increasing in z around z = a. It follows that the same is true of the slightly perturbed
function
θzθ−1 − C(x̂, ŷ)zθ
around z = a. Using this information in (9.23) and (9.24), it is therefore easy to see that for
> 0 but small enough,
S1 (x̂, ŷ) > S2 (x̂, ŷ).
Applying this inequality to (9.22), we may conclude that S(a + , a − ) > S(a, a): surplus is
increased by a small transfer between two persons with the same lower share, and the proof
is complete.
Tentative Notes on Simulations:
Unfortunately, we could not obtain analytically the parameter values for which the unequal
pair dominates equal minority with m = 1, 2. We therefore turned towards simulations, the
results of which we now examine. The following observations were obtained through the
simulations:
(i) The area in the parameter space (α, σ) such that the inefficiency-minimizing distribution
of shares corresponds to the equal minority case is large. In other words, for most values
of the parameters, sharing equally among a limited (smaller than n) number of agents is
the least inefficient distribution of shares. When σ ≥ 12 , perfect equality is the (second-best)
optimal distribution of shares.
(ii) There is a non-trivial area in the parameter space (α, σ) such that inefficiency is minimized
by distributing shares unequally, and between only two agents. However, the two shares
tend to be more equal as the degree of complementarity between efforts increases.
(iii) For σ < 0.2179, the optimal distribution of shares corresponds to the one of perfect
inequality, where one agent concentrates all the shares, for all values of α. Thus, if
substitutability between efforts is high enough, perfect inequality a la Olson appears
to minimize inefficiency. This is a subtle observation that requires greater analytical
investigation.
3The reason for this is that it is possible to take x̂ ∈ (a, a + ) and ŷ ∈ (a − , a). While the statement of the theorem
in Hoffman (1975) does not make this clear, the proof — see last two lines — does.
9.3 Egalitarianism and Incentives
Note. What follows has been taken from Ray and Ueda (1996) and needs to be shortened for
the purpose of these lecture notes.
A social planner’s concern for egalitarianism might lead to a dilution of incentives, and
therefore a loss in efficiency. In an important class of situations, the efficiency loss arises
because the planner cannot credibly commit to a future course of action, such as the
decision not to tax an individual or group making efficiency-enhancing investments. So
these investments are not made, or more generally, undersupplied.
Consider, then, the following class of situations. A group of agents is collectively engaged in
a joint production activity, where the output from production is to be distributed among the
members of the group. The agents (represented by a social planner, perhaps) are interested
in maximizing the value of a Bergson-Samuelson social welfare function defined on their
own utilities. However, while this welfare function represents their social values, individual
actions are taken on an entirely selfish basis. Suppose that to achieve the desired outcome,
each agent must take an observable action, followed by some collective action — the “social
planner’s move”. Suppose, moreover, that a collective action (contingent on individual
decisions) cannot be credibly committed in advance.4
Specifically, define a soft mechanism to be one that specifies a second best division of the
output (relative to the social welfare function) conditional on every possible input vector. We
wish to compare the resulting equilibria of the “soft game” so induced, with the first best
under the very same welfare function.
9.3.1 Model.
9.3.1.1 Technology and Individual Preferences Consider a group of n individuals (n ≥ 2)

producing a single output. Output is produced by the joint efforts of these individuals
according to the production function
(9.25) Y = F(e)
where e = (e1 , . . . , en ) is a (nonnegative) vector of efforts. We assume

∂F(e)
(A.1) F is continuous and concave with F(0) = 0 and differentiable whenever F > 0, with ∂ei
≡
Fi (e) > 0.
Each individual has preferences over pairs of consumption (c) and leisure (l). We assume
4A central example is the problem of allocating funds from a Central Government to different State Governments.
Typically, each state carries out a number of different expenditure programs, which are financed from central
funds and state revenues. The Center would like to channel proportionately greater funds to the poorer states.
At the same time, the Center would like to induce each state to carry out activities that will raise per capita income
in that state. But the former goal places limits on the punishments that a Center can credibly impose on a state
for not taking actions to further certain desired goals.
(A.2) For each i, preferences are representable by an increasing utility function ui : R2+ → R which is
C2 and strictly quasiconcave, with (uic (c, l), uij (c, l)) 0 whenever u(c, l) > u(0, 0). Moreover, leisure
is a normal good.5
Each individual has a labour endowment Li . Thus, for each i,
(9.26) L i = li + e i
Let E be the product of intervals [0, Li ] over all i. Note that whenever e ∈ E, we shall
have li given by (2), for i = 1, . . . , n. Keeping this in mind, define an outcome as a pair
(c, e) = ((c1 , . . . , cn ); (e1 , . . . , en )), such that c ≥ 0, e ∈ E, and
(9.27) c1 + c2 + . . . + cn = F(e)
9.3.1.2 Social Preferences The social planner is presumed to possess Bergson-Samuelson

preferences satisfying standard restrictions:
(A.3) (i) Society’s preferences are representable by a C2 welfare indicator W : R2n
+ → R, defined on
2n-tuples of consumption-leisure vectors. This welfare function is strictly quasiconcave. Moreover,
for each l ∈ E and each i, ci is not an inferior good under the function W(., l).
(ii) There exists a function V (which will be C2 by part (i) and (A.2)) such that W(c, l) =
∂V(u)
V(u1 (c1 , l1 ), . . . , un (cn , ln )). Moreover, ∂ui ≡ Vi (u) > 0 for all i.
9.3.1.3 First Best The planner would like to maximize social welfare. This is achieved by
an outcome (c, e) that maximizes W(c, L − e), where L ≡ (L1 , L2 , . . . , Ln ) is the vector of labor
endowments. From Assumptions 1 and 3 it follows that there is a unique outcome (c∗ , e∗ )
that solves this problem. Call this outcome the first best.
Some minor restrictions on the first-best are summarized in the following assumption.6
(FB) There is at least one individual i such that e∗i > 0 and l∗i ≡ Li − e∗i > 0. Moreover, the first best
output is larger than the optimal output that would be chosen by any one individual acting on his
own: F(e∗ ) > F(0, . . . , 0, ê j , 0, . . . , 0), where ê j solves
max0≤e j ≤L j u j (F(0, . . . , 0, e j , 0, . . . , 0), L j − e j ). Finally, under the first-best, each individual obtains
at least as much utility as total inaction (which yields ui (0, Li )).
9.3.1.4 The Soft Mechanism The idea of a no-commitment mechanism, or soft mechanism,
is based on the postulate that the planner cannot avoid maximizing welfare ex post, even
though this may be detrimental to the maximization of welfare ex ante. Formally, for each
given vector e ∈ E, consider the problem:
(9.28) max W(c, L − e)
c
subject to the constraint that (c, e) must be an outcome.
5The assumption that leisure is a normal good implies the restriction ui ui − ui ui > 0.
c cl l cc
6All the restrictions are not used in all the propositions, but the conditions are so minor that we do not feel it
useful to separate this list into separate components.
By (A.1) and (A.3), there is a unique vector c(e) that solves (4). We will refer to (c(e), e) as an ex
post outcome. This captures, in an extreme way, the inability to impose arbitrary punishments
on deviants. The lack of commitment creates limitations on the incentive mechanisms that
can be used to stimulate production.
The collection of ex post outcomes (c(e), e), for e ∈ E, induces a soft game in the obvious way:
Player i chooses ei ∈ [0, Li ]. If the vector e is chosen, i consumes ci (e), thus generating the
payoff ui (ci (e), Li − ei ).
9.3.1.5 An Additional Restriction We now have sufficient terminology to introduce an

additional joint assumption on technology and preferences, which will be used to derive
one of the main results. To motivate this assumption, first consider a standard property of
the first best outcome, which can be easily verified: If e∗i < Li ,
(9.29) uic (c∗i , l∗i )Fi (e∗ ) ≤ uil (c∗i , l∗i )

(indeed, with equality holding if e∗i > 0).
The intuition is simple. If this inequality did not hold, i’s effort could be raised a little with
all the additional output being credited to him. His utility would be higher, with every other
utility remaining constant. Social welfare goes up, a contradiction to the fact that we have a
first best outcome to start with.
The additional assumption that we wish to make is related closely to (5). Specifically, we
suppose:
(A.4) For each ex post outcome (c(e), e) such that F(e) ≥ F(e∗ ), there is i with ei > 0, and
(9.30) uic (ci , li )Fi (e) ≤ uil (ci , li )
where ci (e) ≡ ci > 0.
This assumption looks plausible, because as output moves above the first best, we would
expect that marginal products do not increase, while the marginal rate of substitution
between consumption and leisure tilts in favor of leisure. Thus if (5) already holds at
the first best, we expect this relationship to be maintained (for at least one individual) for
outcomes with higher output.
However, it is only fair to point out (A.4) is not automatically implied by (A.1) - (A.3).
However, experimentation with different functional forms suggests that it is implied by a
large subclass of welfare functions, individual preferences and production technologies.
For instance, suppose (in addition to (A.1)–(A.3)) that individual utilities are separable as
the sum of concave functions of consumption and leisure, that total output is some concave,
smooth function of the sum of individual efforts, and that the social welfare function has a
separable and concave representation in utilities.
Consider some ex post outcome (c(e), e), distinct from the first best, but with the property
that at least as much output is being produced as in the first best: F(e) ≥ F(e∗ ). There must
be some individual i with ei > e∗i . In the case under consideration, it will be the case that
at least one such individual gets ci (e) ≥ c∗i (remember that at least as much output is being
produced). Furthermore, since output is no lower, marginal product cannnot have increased
relative to the first-best. Putting all this together with (5) and using the convexity properties
of preferences and technology, it should be the case that
uic (ci (e), Li − ei )Fi (e) ≤ uil (ci (e), Li − ei )
for this individual, which is (6).
Consider, therefore, the following formalization of the separable case:
(A.4*) (i) ui is additively separable in (c, l), (ii) F(e) is of the form f (e1 + . . . + en ), for some
f : R+ → R+ , and (iii) V is separable in individual utilities.
Our motivating argument above shows that under (A.1)—(A.3), (A.4*) implies (A.4).
9.3.2 Egalitarianism and Incentives.
9.3.2.1 Egalitarianism Yields Underproduction...

Proposition 9.10. Under (A.1) - (A.3) and (FB), the first best cannot be achieved as some equilibrium
of the soft game. Moreover, if (A.4) holds, every equilibrium of the soft game must involve
underproduction.
Proposition 9.10 verifies an intuitive conjecture: when a team cannot precommit to

adequately punish deviants, the first best cannot be achieved. Specifically, in this model
there is underproduction relative to the first best.
It is perhaps worth mentioning that while Proposition ?? appears intuitive (especially
because it is related to other inefficiency notions, such as the Marshallian inefficiency
of sharecropping, or the holdup problem), it is far from being obviously true in the
present context. For instance, if individual utility has a separable linear representation
in consumption, the proposition is, in general, false: there are equilibria that attain the first
best.7 Moreover, without (A.4), whether there is underproduction or not is an open question.
Most significantly, the assumptions of Proposition 9.10 do not cover the case of Rawlsian
social preferences, and as we shall see in the next section, our proposition fails in this case.
It might help, therefore, to sketch the proof of Proposition 9.10 for a special case. Assume
that individual preferences have a strictly concave representation that is separable in
consumption and leisure, and that given this representation, the Bergson-Samuelson welfare
function has a representation that is strictly concave and separable in individual utilities.
Furthermore, assume that total output is a concave, increasing function f of the sum of
individual efforts. This is the case covered by (A.4*).
Now consider an ex post outcome (c, e) with overproduction relative to the first best. Someone
must be putting in more effort relative to the first best. Call that someone i. Evaluated at
the first best consumption vector, i must now have a higher marginal weight in the social
welfare function than he had before. In the ex post outcome under consideration, therefore, i
7Such utility functions are ruled out by the assumption that leisure is a normal good (see (A.2)).
must enjoy more consumption, because total output is not lower. Putting all this information
together with the convexity of preferences and technology, it is easy to see that
(9.31) uic f 0 (e1 + . . . + en ) ≤ uil
which just means that (A.4) is automatically satisfied in this special case.
Now, suppose that i reduces his effort level a tiny bit. Perform the thought experiment of
cutting i’s consumption by exactly the resulting fall in output. By the inequality (9.31), this
has a nonnegative first-order effect on i0 s utility. If the outcome is to be a Nash equilibrium,
therefore, i’s true consumption decline must be at least of this order. But now note that this
decline must have a first-order effect on i’s marginal utility of consumption, which rises.
Moreover, by the supposition that i is no better off, the marginal welfare contribution of his
utility is no lower. It is now easy to check that in this new situation, the first-order conditions
for ex post welfare maximization are destroyed. For because i is being made to take at least
the entire output loss (as we have argued above), there must be some other individual whose
consumption is no lower. For this other person, exactly the reverse changes must occur in
the marginal conditions. So, if the first order conditions were holding earlier (as they must
have), they cannot be holding now! This contradicts ex post welfare maximization, and
proves that the original outcome could not have been an equilibrium.
9.3.2.2 ...But Not With a Rawlsian Planner Our main theme relates the degree of egalitari-
anism to the degree of efficiency failure. It will be convenient to begin by exploring the most
extreme form of egalitarianism: Rawlsian social preferences. In a later section, we will carry
out the more complicated exercise of varying the welfare function over different degrees of
egalitarianism.
Observe that the Rawlsian case is not covered by Proposition 9.10. That result rests on a
postulate that is seemingly so innocuous that we have not emphasized it in the discussion
(though, of course, it is formally stated in (A.3)). It is that social welfare is strictly increasing
in every utility level, for each vector of utilities. Rawlsian social preferences do not satisfy
this condition. Indeed, this observation has striking consequences, as we shall see.
We first define Rawlsian preferences. To do so we need a benchmark comparison. This is
summarized in the following assumption:
(R) Society is indifferent between the complete inaction of any two individuals. Thus if the Rawlsian
welfare is written as
(9.32) V(u) = min ui
i
where utilities have already been normalized so that ui (0, Li ) = u j (0, L j ) for all i and j.
The first best Rawlsian optimum involves the maximization of the expression in (9.32),
subject to the constraints that ui = ui (ci , li ) for all i and that (c, e) is an outcome. Just as before,
there is a unique first best outcome; call it (c∗ , e∗ ).
Ex post outcomes are defined exactly as they were earlier. It is easy to see that for each e ∈ E,
there is a unique consumption vector c(e) that solves the ex post maximization problem.
The collection of ex post outcomes induces a soft game just as before, and we may study its
equilibria.
Proposition 9.11. Under (A.1), (A.2), (FB) and (R), the Rawlsian first best is an equilibrium of
the Rawlsian soft game. Furthermore, any equilibrium which gives at least one agent strictly more
utility than the utility from inaction must be the Rawlsian first best outcome.
(1) One might wonder whether the strikingly positive result of Proposition 9.11 is due to the
fact that the Rawlsian egalitarian optimum is so devoid of efficiency properties that we do
not have any incentive problem at all to maintain it. But this is not the case. For instance,
consider any special case of this model that is symmetric across all agents. It is easy to see
that the first best outcome is invariant across all symmetric quasi-concave welfare functions,
including the Rawlsian one.
(2) There is only one qualification in Proposition 9.11. To prove that an equilibrium must be
the Rawlsian first-best, we assume that at least one individual receives strictly more utility
than he receives from inaction. There may exist an equilibrium involving total inaction. This
will happen if the technology has the property that F(e) = 0 whenever n − 1 components of e
equal zero. On the other hand, if the technology is such that output depends on the sum of
the efforts (and if right-hand marginal utilities are defined and strictly positive everywhere)
then our qualification can be dispensed with.8
Thus, in contrast to standard intuition, extreme egalitarianism might actually has pleasing
incentive properties. Egalitarianism applies not only to the choice of the social optimum,
but in the treatment of deviants from the optimum. A greater concern for egalitarianism
goes hand in hand with the ability to credibly mete out stronger punishments.
Apart from the technicalities, the proof of Proposition 9.11 is very simple and general.
Under convexity of the feasible set and preferences, the Rawlsian criterion has the following
property. All individual utilities move in the same direction from one ex post outcome to
another. Consequently, the Rawlsian first best is always an equilibrium of the Rawlsian soft
game. For if someone could improve his utility by a deviation, he would improve the utility
of everyone else in the process. This would contradict the fact that the earlier outcome was
first best.9
The second part of the result — that every equilibrium must be first best — is model-specific
in two respects. The convexity and the differentiability features of the model must both be
exploited. These are used to guarantee that if an allocation is not first best, then there is
some small, unilateral change in someone’s effort level that creates an ex post outcome with
8Our requirement that the production function be differentiable at positive output is used in the proof. Concave,
increasing production functions which produce no output when n − 1 effort components equal zero are not (right
hand) differentiable at 0, even though all partial (right hand) derivatives may exist at 0.
9
It should be pointed out that this joint movement of utilities in the same direction is also a feature of the case in
which utilities have a separable linear representation in consumption. To see this, let ui (c, l) = ai c + bi (l), where
ai > 0 and bi is an increasing, smooth, concave function. Assume that V(u) = v1 (u1 ) + · · · + vn (un ), where each vi
is smooth, increasing, and strictly concave. Now look at the ex post outcomes of this model. If all consumptions
are strictly positive, then the necessary and sufficient conditions characterizing ex post optimality are that for
all i, j, v0i (ui (ci , li ))ai = v0j (u j (c j , l j ))ai . The co-movement of utilities should now be apparent. So in this case as well,
the first best can be supported as an equilibrium of the soft game.
a higher Rawlsian value. Again, using the Rawlsian criterion and the equal-utility property
yielded by convexity, the individual who makes the change must participate in its benefits,
thereby destroying the equilibrium possibilities of the given outcome.
9.3.2.3 Changing Egalitarianism The result of the previous section, and the discussion
following it, suggest an even stronger observation: that as the extent of egalitarianism
increases, the degree of underproduction should monotonically decline. Proposition 9.11
would be the limiting case of such an observation. The purpose of this section is to
demonstrate such a possibility.
We consider a symmetric version of our model, described formally as follows. First, any
permutation of e produces the same output level as F(e). Second, every individual’s
preferences is represented by the same function u. Finally, the social welfare function is
symmetric:
(9.33) W(c, l) = V (u(c1 , l1 ), ..., u(cn , ln ))

where V is symmetric. We will fix the individual representation u, and analyze the effect of
changing egalitarianism by altering the form of V in a manner made precise below.
The following assumption (in addition to those already maintained) will be made on the
fixed cardinal representation, u, of individual preferences.
(A.5) There exists a cardinal representation such that u is strictly concave in c, and V is quasiconcave
in u.10
We begin by discussing how to compare the “degree of egalitarianism” among different

V’s, or more generally, among different social welfare orderings on vectors of individual
utility. Our definition, while not formally requiring symmetric social welfare orderings, is
0
best viewed in this background. Let S and S be two social welfare orderings for utility
0
vectors u. We will say that S is at least as egalitarian as S if for all u, v ∈ Rn , and for all i and
j, the condition
(9.34) uk = vk for k , i, j, vSu, and |vi − v j | < |ui − u j |
implies vS0 u.
0
That is, S is more egalitarian than S if it prefers every redistribution plan to narrow the
difference between two individuals’ utility that R prefers as well . This can be viewed as an
extension of the Pigou-Dalton principle for utility distributions (see, for example, Moulin
[13]), applied to a comparison of any two social welfare orderings.
Some remarks on this last observation will clarify our definition. Observe that under our
definition, a social welfare ordering S0 is at least as egalitarian as the utilitarian welfare
ordering if the condition
uk = vk for k , i, j, ui + u j ≤ vi + v j , and |vi − v j | < |ui − u j |
10The fix on a cardinal representation that is strictly concave is not necessary for the results (as long as (A.3) is
maintained), but it greatly simplifies the writing of the proofs.
implies vS0 u. Note that the condition above is just a specialization of (9.34), when S is
the utilitarian ordering. Specializing even further so that ui + u j = vi + v j in the above
condition, we obtain the Pigou-Dalton principle for utility distributions. In other words, a
social welfare ordering which is at least as egalitarian as the utilitarian ordering satisfies the
Pigou-Dalton principle. Our suggested relation is therefore an extension of this idea.
The argument above suggests that as we progressively move “towards egalitarianism” in
this ordering, we obtain social welfare functions that are “more willing” to trade off total
utility for interpersonal equality.
This definition, and the subsequent discussion leads to two general observations. First,
in the class of social welfare functions considered in this paper, every such function is at
least as egalitarian as the utilitarian function.11 Second, the Rawlsian function is at least
as egalitarian as any of these social welfare functions.12 In short, the utilitarian and the
Rawlsian social welfare orderings are two extremes of the social welfare functions we are
considering, in the spectrum of the degree of egalitarianism defined here.
In our model, social welfare orderings are given by the functions satisfying (A.2) and (A.3),
0
which includes differentiability. Let V and V be two such symmetric functions. Then we can
0
show that V is at least as egalitarian as V if and only if
0
Vi Vi
(9.35) For all u and for all i, j ∈ {1, ..., n}, ui ≥ u j ⇔ 0 ≤
Vj Vj
We omit a proof of this result (it is available on request).13 But the interpretation should
be quite natural. When one individual j has a lower utility than another individual i, the
“more egalitarian” social welfare function assigns a higher (relative) marginal welfare to i.
The social marginal rate of substitution between i and j is tilted in favor of j by the more
egalitarian welfare function.
To obtain a clearer idea of the partial ordering proposed here, consider two classes of social
welfare functions. The first is the Atkinson family, which is represented by the form
n
ρ
X
(9.36) [ ui ]1/ρ
i=1
11This is because increasing, strictly quasiconcave welfare functions will indeed rank v over u whenever the
specialization of condition (9.34) discussed in the text happens to hold.
12
This is easily seen as follows. Suppose that for some social welfare ordering S, and utility vectors u and v,
condition (9.34) holds but nevertheless, u is preferred by the Rawlsian ordering to v. This means, in particular,
that min{ui , u j } > min{vi , v j }. But because |ui − u j | > |vi − v j |, this implies that (ui , u j ) vector dominates (vi , v j ). But
now we have a contradiction to the assumption that vSu.
13It is worth noting that the asserted equivalence is an analogue of a result in Moulin [13, Lemma 2.2], which
characterizes a differentiable social welfare function (called a “collective utility function”) that satisfies the
Pigou-Dalton principle for utility distributions.
where ρ ∈ (−∞, 1]. It is easy to check that higher degrees of egalitarianism correspond
to lower values of ρ. In particular, the case ρ = 1 yields the utilitarian, and ρ = −∞, the
Rawlsian social welfare ordering.
The second class is the ”constant absolute risk-aversion” family, which is given by the form
n
X
(9.37) − exp(−ηui )
i=1
where η ∈ (0, ∞). A higher value of η is associated with a more egalitarian function, and
the two extreme values of η (0 and ∞) again give the utilitarian and the Rawlsian social
orderings respectively.
Now, let us return to the model. An outcome is symmetric if all individuals work equally hard,
and consume equally. In the symmetric model, symmetric effort by individuals induces a
symmetric outcome, in which c(e, ..., e) = (c, ..., c) and c = n1 F(e, ..., e). The first best (c∗ , e∗ ) is
also a symmetric allocation, and the allocation is independent of the social welfare function.
We will restrict our discussion to the class of symmetric equilibria of the soft game.
In what follows, it will be convenient to consider a modified soft game where all but the first
individual choose the same level of effort. Then the ex-post outcome assigns the same level
of consumption for those providing this effort. Denote the generated levels of consumption
by
(9.38) (c1 (e1 , e; V), c(e1 , e; V))

where V is the social welfare function determining consumption, and e and c are the common
values for all but the first individual. The first individual’s best response to e is the set of
solutions to the problem of maximizing his indirect utility function with respect to e1 ∈
[0, L]. Denote this set by B(e; V). B(e; V) is a compact-valued and upper hemi-continuous
correspondence of e . We will need to further assume:
(A.6) u(c1 (e1 , e; V), L − e1 ) is single-peaked with respect to e1 .
Then, B(e; V) is convex-valued and the existence of a symmetric equilibrium for each V is
guaranteed.14
By the degree of underproduction in any outcome, we refer to the difference between the
first-best production level and the level of aggregate production under that outcome. The
following proposition establishes that while the degree of underproduction is positive, the
degree of underproduction must fall with rising egalitarianism.
Proposition 9.12. Assume a symmetric model satisfying (A.1)—(A.3), (A.5), (A.6) and (FB). Then
every symmetric equilibrium of the soft game involves underproduction. Now consider two welfare
14It should be noted that our maintained assumptions do not automatically yield (A.6). However, all parametric
forms that we have tried satisfy this assumption.
0 0
functions V and V , where V at least as egalitarian as V. Then, for each symmetric equilibrium
0
under V, there exists a symmetric equilibrium under V such that the degree of underproduction is
not more for the latter.
This verifies that the Rawlsian case discussed in Proposition 9.11 is not an exception or some
quirky failure of continuity. For symmetric equilibria of the symmetric model, increased
egalitarianism never increases the degree of underproduction.15
9.3.3 Extensions. We briefly consider some extensions, as well as possible objections to

the setting in which these results have been derived.
1. Sinking past differences: In computing ex post optima, why do teams take account
of the sunk efforts already incurred by its members? It may be argued that bygones are
bygones, and that the social welfare function should ignore this. This point of view may be
identified with the assertion that individual utilities “should” be separable in consumption
and leisure, and that the social welfare function “should” be utilitarian. At least, that is
the only way to justify the assertion if one sticks to the Bergson-Samuelson setting. Such
particular functional forms are already accommodated as a special case of the paper, and
if one insists on such an interpretation, an entire class of welfare functions (including the
Rawlsian function) is simply removed from consideration.
On the other hand, one might take the sinking of past differences as a primitive, simply
asserting that ex post output is shared equally (in a symmetric model) irrespective of the
welfare function. We find it difficult to see what might justify such an assumption. Consider
any resource allocation problem. As long as resources are not allocated to everybody at
exactly the same point of time, this point of view leads to absurd allocations.
If one were to admit such a structure for the sake of argument, inefficiencies will always
arise, of course. But even then, it can be shown that as long as some weight is given to the
past, Proposition 9.12 will continue to hold. In particular, the Rawlsian welfare function will
exhibit the lowest degree of underproduction.
2. Emotions and credible punishments: One might proceed in the exactly the opposite
direction to (1). Deviants that do not adhere to the desired outcome might enjoy an entirely
different (and reduced) weight in the ex post welfare function. Emotions such as anger
or social disapproval might induce such changes, and in so doing, lend credibility to
punishments (for similar ideas, see, e.g., Frank [6]). It is to be expected that the degree
of inefficiency will be lowered in the presence of these emotions. But even so, as long as
the first best is thereby not automatically achieved, more egalitarian welfare functions will
possess better efficiency properties.
3. The observability of effort: The results of this paper rest critically on the assumption
that efforts are observable. We suspect that the results would extend to noisy observability,
15It is possible to find even stronger versions of this proposition in special cases. For instance, if the best-response
correspondence is single-valued, and social welfare functions are in the Atkinson class, we can show that
increased egalitarianism strictly lowers the maximal degree of underproduction. The qualification “maximal”
is due to the fact that equilibrium may not be unique.
though we have not checked this. Of course, if efforts are not observable at all, the results
cease to have any relevance.
4. Repeated Relationships: In a dynamic situation, teams would recognize that a departure

from the soft mechanism will serve them well in the longer run. It may be of interest to
study such repeated relationships. But as a prelude to that study, it is surely important to
analyze the “one-shot” relationship without precommitment, which is exactly what we do
here.
9.3.4 Discussion. While the connections between egalitarianism and inefficiency has long
been a subject of debate, there have been surprisingly few attempts to model the exact nature
of the tradeoff. To be sure, the sources of various tradeoffs are manifold in nature. This paper
investigates one potential source, and shows that the commonly held intuition is not valid,
at least in this case.
Specifically, this paper studies the idea that egalitarianism fails to uphold proper incentives
because credible punishments are thereby destroyed. This statement really has two parts to
it: one is the familiar “dilemma of the Samaritan” induced by the inability to precommit.
Organizations that cannot precommit, yet derive their sense of goal-fulfillment or welfare from
the welfare of its members, are particularly prone to these potential inefficiencies. Indeed,
organizational structures where the source of utility for the “principal” is directly opposed
to that for the “agent” will certainly do better (in terms of efficiency) compared to the
structures considered here.16 This is the intuition upheld by Proposition 9.10. An inability to
precommit the reward function results in inefficiency. While the particular result proved is, to
our knowledge, new, there is nothing particularly surprising or novel about the underlying
theme.
But this is only one half of the idea. The second part goes further. It states that egalitarian
organizations faced with the inability to precommit are doubly cursed: they are inefficient on
the additional count that they cripple the incentive system (already weakened by the inability
to precommit) even further. In this paper, we argue that this assertion is wrong. Increased
egalitarianism restores incentives that are damaged by the lack of commitment (Proposition
9.12). Indeed, in the extreme case of Rawlsian egalitarianism, the precommitment and
no-precommitment yield exactly the same first-best outcomes.
The results in this section are provocative on two counts. First, they might inspire greater
interest in a challenging and crucially important area of research: the connections between
the ethic of equality and the yardstick of aggregate performance (such as GNP growth). The
second aspect is one that we have not emphasized in this paper, but of great interest, we
believe. This is the connection with the theory of implementation. By far the dominant
16Indeed, in the usual principal-agent setup the problem is trivial. If efforts are observable, the organization
is best off using a forcing contract. The contract demands that each individual must supply (at least) his first
best effort level. If this effort level is indeed supplied by all individuals, all individuals receive the first best
output division. If, however, these effort levels are not forthcoming, the contract promises dire consequences for
the individuals who have deviated. These threats indeed implement the first best (at least as one equilibrium),
and there is no incentive problem worth serious analysis. Authors such as Holmstrom (Holmstrom [8]) have
argued that the existence of a residual claimant (capitalist, manager) creates a credible threat to carry out these
punishments, such as retention of produced output, in the event of breach.
approach in this literature presumes that the planner can implement outcomes without
regard to ex post credibility. But there are many situations where it is natural to constrain
mechanisms off the equilibrium path by the provision that they should not be suboptimal
relative to the actions that have been taken and observed, and the social welfare function of
the planner. This may have some bearing on narrowing the class of allowable mechanisms
in implementation contexts.17
9.3.5 Proofs. Note. Equation numbering needs to be changed.

Lemma 9.2. Consider some ex post outcome (c(e), e) such that F(e) ≥ F(e∗ ), with the additional
property that for some individual i,
(9.39) uic (ci , li )Fi (e) ≤ uil (ci , li )
Then such an outcome cannot be an equilibrium of the soft game.
Proof. Suppose that (c(e), e) ≡ (c, e) is an ex post outcome with F(e) > 0. Denote by Wi (c, l)
the partial derivative of W with respect to ci . Then, under (A.1)–(A.3) and using the ex post
maximization problem (4), (c, e) must satisfy the following property: if ck > 0, then for all
j = 1, . . . , n,
(9.40) Wk (c, l) ≥ W j (c, l)
with equality whenever c j > 0.
We need a slightly stronger implication than (16). Let i be such that ei > 0, and suppose
that there is some k with ck = 0. Suppose, further, that there exists > 0 such that for all
e0i ∈ (ei − , ei ), the ex post consumption vector c(e0i , e−i ) ≡ c0 has c0k > 0. Then, indeed, we
can say that (16) holds for this k even if ck = 0. We exclude the verification of this simple
observation.
Now, suppose that, contrary to the lemma, this outcome is an equilibrium of the soft game.
Let M be the set of all indices j such that either c j > 0, or with the property that there is > 0
such that for all e0i ∈ (ei − , ei ), the ex post consumption vector c(e0i , e−i ) ≡ c0 has c0j > 0. We are
going to consider the effect (on i’s utility) of a small reduction in ei by differential methods.
By the maximum theorem (and the uniqueness of ex post consumption (given effort)), c(e)
is a continuous function. Therefore the set M is all that counts for the analysis, and for all
k, j ∈ M, (16) holds with equality.
Note first that because (c, e) is an equilibrium and ei > 0, we have ci > 0. Consequently, i ∈ M.
Without loss of generality, number the indices in M as 1, . . . , m, and let i be rechristened with
the index 1. We now claim that m ≥ 2. Suppose not. Then, because (c, e) is an equilibrium, we
must have e j = 0 for all j , 1. Therefore, because F(e) ≥ F(e∗ ) and because we have assumed
that the first best output is larger than the output of any individual acting completely on his
own, we must have
(9.41) u1c (c1 , l1 )F1 (e) < u1l (c1 , l1 )
17Maskin and Moore [11] study a related problem where the ability of the planner is limited by the possibility
that agents may renegotiate the outcome.
Because M = {1}, it follows from (17) and the definition of M that a small reduction in e1 will
raise 1’s welfare, because all other consumptions will continue at zero. This contradicts our
supposition that (c, e) is an equilibrium, and shows that m ≥ 2.
For all i ∈ M, we have the first order condition of the ex post maximization problem: for
some λ < 0,
Wi (c, l) + λ = 0
(9.42) c1 + . . . + cm = F(e)
We are interested in differentiating this system with respect to a parametric change in e1 ,

and studying dc 1 18
de1 . Given our remarks above, the differentiation argument will reflect the
true story for small changes in e1 , provided that we think of these changes as reductions.
∂2 W
Differentiating (18) and defining Wi j ≡ ∂ci ∂c j
, we obtain the system
  F1 (e)de1
 
0 1 1 ... 1   dλ   ∂W1 (c,l)
  
de1
 
 1 W11 W12 ... W1m   dc1   ∂li

  dc2   ∂W2 (c,l) de1
 
...
   
(9.43)

 1 W21 W22 W2m = ∂li


.. .. .. .. ..   ..   ..
   

 . . . . .   .  
    .


1 Wm1 Wm2 . . . Wmm  ∂Wm (c,l)
 
dcm de1

∂li
∂Wi (c,l)
Let us calculate ∂li
. For i = 1,
∂W1 (c, l) ∂[V1 (u)u1c (c1 , l1 )]

(9.44) = = V11 (u)u1c u1l + V1 (u)u1cl
∂l1 ∂l1
For i , 1, we have
∂Wi (c, l) ∂[Vi (u)uic (ci , li )]

(9.45) = = Vi1 (u)uic u1l
∂l1 ∂l1
Next, let us calculate Wi1 (c, l). For i = 1,
∂[V1 (u)u1c (c1 , l1 )]

(9.46) W11 (c, l) = = V11 (u)[u1c ]2 + V1 (u)u1cc
∂c1
while for i , 1, we see that
∂[Vi (u)uic (ci , li )]

(9.47) Wi1 (c, l) = = Vi1 (u)u1c uic
∂c1
18(9.42) defines (c, λ) implicitly as a function of the “parameter” e. Because output is positive, it follows from
(A.1)–(A.3) that this function is C1 . Moreover, because the bordered Hessian in (9.43) has non-zero determinant
(see footnote 6), it follows from the implicit function theorem that c is a differentiable function of e, and in
particular of e1 .
Now, recalling that leisure is a normal good in individual preferences, we have u1c u1cl −u1l u1cc >
0 (recall (A.2) and footnote 7). Using this and manipulating (20), we have
∂W1 (c, l) u1l u1c u1cl
= (V11 (u)[u1c ]2 + V1 (u) )
∂l1 uc1 u1l
u1l
> (V11 (u)[u1c ]2 + V1 (u)u1cc )
u1c
u1l
(9.48) = W11 (c, l)
u1c
while from (21) and (23) it is easy to see that for all i , 1,
1
∂Wi (c, l) ul
(9.49) = 1 Wi1 (c, l)
∂l1 uc
Now, let us return to (19) and write down the solution for dc 1
de1 . Let U be the determinant
of the cofactor of the first 1 (from the left) in the first row of the matrix in (19), and Ui be
the determinant of the cofactor of Wi1 , i = 1, . . . , m. Include in these the signs generated by
cofactor expansion. Let U∗ be the determinant of the matrix in (19). Then it can be seen that
dc1 1 ∂W1 (c, l) ∂W2 (c, l) ∂Wm (c, l)
= [F1 (e)U + U1 + U2 + . . . + Um ]
de1 U∗ ∂l1 ∂l1 ∂l1
Recalling from the strict quasiconcavity of W(.) and (18) that U1 and U∗ have different signs,
and using (24) and (25), we have
dc1 1 u1l
(9.50) < [F1 (e)U + 1 (W11 U1 + W21 U2 + . . . + Wm1 Um )]
de1 U∗ uc
Next, because c1 is a noninferior good under W(., l), we have U/U∗ ≥ 0. Consequently,
combining (9.39) and (26),
dc1 u1l 1
< [U + W11 U1 + W21 U2 + . . . + Wm1 Um ]
de1 u1c U∗
u1l
(9.51) =
u1c
Therefore, using (27),
du1 (c1 , l1 ) dc1
(9.52) = u1c − u1l < 0
de1 de1
so that (28) proves that a small reduction in e1 will raise the utility of 1 under the new ex
post outcome. This contradicts our supposition that (c, e) is an equilibrium, and completes
the proof of the lemma.
Proof of Proposition 1. Observe that the first best allocation is indeed an ex post outcome,
with positive output, satisfying (9.39) in Lemma 1. Given the assumption that at the first
best, there is i with (e∗i , l∗i ) 0, (9.39) follows as a necessary condition describing the first
best. By Lemma 1, this outcome cannot be an equilibrium of the soft game.
If (A.4) holds, then consider any ex post outcome with F(e) ≥ F(e∗ ). It follows right away
from (A.4) that (9.39) is satisfied. By Lemma 1, this outcome cannot be an equilibrium of the
soft game.
Proof of Proposition 2. For any outcome (c, e), define its Rawlsian value as R(c, e) ≡
mini ui (ci , li ). Now note that under the first best outcome, we must have, for every i,
(9.53) R(c∗ , e∗ ) = ui (c∗i , l∗i )
The reason is that for each i, either c∗i > 0 or l∗i > 0 (or both), because ui (c∗i , l∗i ) ≥ ui (0, Li ).
Therefore, if (29) were not true, we could always improve the Rawlsian value of the outcome
by small changes.
Given this claim, we first prove that the first best outcome is indeed an equilibrium. Suppose
not. Then for some i, there is ei ∈ [0, Li ] such that
(9.54) ui (ci , li ) > ui (c∗i , l∗i )
where ci is the ith component of c ≡ c(e), where e ≡ (ei , e∗−i ).
Because ui (c∗i , l∗i ) ≥ ui (0, Li ) and because (30) holds, it must be the case that ci > 0. It follows
that
(9.55) R(c, e) = ui (ci , li )
For if not, the Rawlsian value of (c, e) can be improved by only changing the consumption
allocation, a contradiction to the fact that (c, e) is an ex post outcome (use the fact that ci > 0).
Combining (29), (30) and (31), we have
R(c, e) > R(c∗ , e∗ )
but this contradicts the fact that (c∗ , e∗ ) is the Rawlsian first best.
Now we prove the second part of the proposition. First, we show that if (c, e) is an
equilibrium, then for all i = 1, . . . , n,
(9.56) R(c, e) = ui (ci , li )
Suppose this is not true. Then for some pair i, j, ui (ci , li ) > u j (c j , l j ). But then it must be the case
(by an earlier argument) that ci = 0. Consequently, u j (c j , l j ) < ui (0, li ) ≤ ui (0, Li ) = u j (0, L j ),
where the last equality follows from normalization. But then, (c, e) cannot be an equilibrium,
for j can guarantee himself at least u j (0, L j ) by deviating.
With this established, we return to the main proof. Suppose that the proposition is false.
Then there exists an equilibrium (c, e) , (c∗ , e∗ ), such that at least one player gets strictly
more utility than inaction. By virtue of (32) and assumption (R), it follows that every player
gets strictly more than inaction. This observation will be used below.
Because the first best is unique, R(c∗ , e∗ ) > R(c, e). Define e(t) ≡ te + (1 − t)e∗ for t ∈ [0, 1],
and c(t) ≡ c(e(t)). Then it is easy to check that R(c(t), e(t)) > R(c, e) for all t ∈ (0, 1) (use (A.1),
(A.2), (29) and (32)). Using (32), it follows that for all t ∈ (0, 1) and all i,
(9.57) ui (ci (t), li (t)) > ui (ci , li )
Now we claim that the following is true:

j j
There exists j such that either (i) e j = L j and uc (c j , l j )F j (e) < ul (c j , l j ), or (ii) e j ∈ (0, L j ) and
j j j j
uc (c j , l j )F j (e) , ul (c j , l j ), or (iii) e j = 0 and uc (c j , l j )F j (e) > ul (c j , l j ).
To prove the claim, first recall that each player gets strictly more than inaction at (c, e), so
that by (A.2), (uic (ci , li ), uil (ci , li )) 0 for all i. Using (A.2) again, together with (33), we see
that for all i,
uil
(ci (t) − ci ) + (li (t) − li ) i > 0
uc
or equivalently, for all i,
uil
(9.58) (ci (t) − ci ) − (ei (t) − ei ) >0
uic
Now, if the claim is false, then, noting that ei (t) ∈ [0, Li ] for all i, we can use (34) to deduce
that
(ci (t) − ci ) − (ei (t) − ei )Fi (e) > 0
Summing this inequality over all i, we have
(9.59) F(e(t)) − F(e) > (e1 (t) − e1 )F1 (e) + . . . + (en (t) − en )Fn (e)
Now observe that because (c, e) provides more utility than inaction, F(e) > 0. But then (35)
contradicts (A.1), which assumes that F is differentiable whenever F(e) > 0, and that F is
concave. This establishes the claim.
Pick j as given by the claim. If part (i) of the claim is true, notice that because (c, e) is an
equilibrium, c j > 0. Then, a small reduction in e j makes j strictly better off even if j pays for
the entire reduction in output from his own consumption. This improvement in j’s utility
persists even if j must pay a small additional amount to each of the other agents. This proves
that the Rawlsian value of (c, e) can be improved by a unilateral change made by j, by going
to an ex post outcome (c0 , e0 ) (where e0 differs from e only in the jth component). That is,
R(c0 , e0 ) > R(c, e). But note that, as a result, u j (c0j , l0j ) ≥ R(c0 , e0 ) > R(c, e) = u j (c j , l j ) (where the
last equality uses (32)). This contradicts the supposition that (c, e) is an equilibrium.
Finally, if parts (ii) or (iii) of the claim are true, use the same argument as above, if
j j j j
uc (c j , l j )F j (e) < ul (c j , l j ), or its reverse (increase e j ), if uc (c j , l j )F j (e) > ul (c j , l j ), to arrive at
a contradiction.
This completes the proof of the proposition.
Proof of Proposition 3. By our assumptions, the first best is symmetric and has
0 < e∗ < L.
It follows from the first-order conditions characterizing the first best that
uc (c∗ , L − e∗ )Fi (e∗ , . . . , e∗ ) = ul (c∗ , L − e∗ ),
and moreover, for every e > e∗ and c ≡ n1 F(e, . . . , e),
uc (c, L − e)Fi (e, . . . , e) ≤ ul (c, L − e).
Moreover, these hold for all i. It follows from Lemma 1 that no symmetric e ≥ e∗ can be an
equilibrium of the soft game.
To prove the remainder of the proposition, fix a common utility level ū for all but the
first individual, and let u1 denote the utility level of individual 1. Then by applying the
characterization of increased egalitarianism for smooth welfare functions, we see that
0
V2 V2
(9.60) ū ≥ u1 ⇔ 0 ≤
V1 V1
where a subscript 1 denotes the marginal social welfare contributed by the first individual’s
utility, and the subscript 2 is the corresponding contribution by the others. Define a “utility
possibility frontier” (given efforts) in the modified soft game by Ψ(ū, e1 , e):
(9.61) Ψ(ū, e1 , e) ≡ u(F(e1 , e, ..., e) − (n − 1)c, L − e1 )

where c is chosen such that ū = u(c, L − e). Ψ is downward sloping and concave with respect
to ū, by (A.5).
The social planner chooses consumption to maximize V on Ψ. Let us introduce some
notation. We will denote by U1 (e1 , V) the utility of agent 1 when the action taken by him is
e1 , the welfare function is V, the action of the remaining agents is fixed at e (which is not
explicitly carried in the notation), and the planner chooses ex-post consumption optimally.
Likewise, we will denote by U(e1 , V) the utility of each of the other agents under exactly the
same state of affairs.19 We make two observations. First,
1
(9.62) F(e, ..., e)
c1 (e, e; V) = c(e, e; V) =
n
for any V. Second, it is possible to show, using the concavity of Ψ, the strict quasiconcavity
of V and V 0 , and (36),20 that
(9.63) U1 (e1 , V) ≥ U(e1 , V) ⇔ U1 (e1 , V) ≥ U1 (e1 , V 0 ) ≥ U(e1 , V 0 ) ≥ U(e1 , V),
when V 0 is at least as egalitarian as V. Moreover, precisely the opposite chain of inequalities
hold on the RHS of (39) if the opposite inequality holds on the LHS of (39).
Let V 0 be at least as egalitarian than V. We claim that
[I] If e ∈ B(e; V), then e ≤ max B(e; V 0 ).

19Thus U (e , V) = u(c (e , e, V), L − e ) and U(e , V) = u(c(e , e, V), L − e).
1 1 1 1 1 1 1
20In fact, this is the only point at which (36) is used in the proof.
Suppose not. Then max B(e; V 0 ) < e. So there exists x ∈ [0, e) such that
(9.64) U1 (x, V 0 ) > U1 (e, V 0 ).
Because e ∈ B(e, V), it must be the case that
(9.65) U1 (x, V) ≤ U1 (e, V)
Noting from (38) that U1 (e, V) = U1 (e, V 0 ), we may combine (40) and (41) to obtain
(9.66) U1 (x, V) < U1 (x, V 0 )
Combine (42) with (39) and the claim immediately following (39). We may deduce that
(9.67) U1 (x, V 0 ) ≤ U(x, V 0 )
Noting from (38) that U1 (e, V 0 ) = U(e, V 0 ), and combining this observation with (40) and (43),
(9.68) U(x, V 0 ) ≥ U1 (x, V 0 ) > U1 (e, V 0 ) = U(e, V 0 ).
In words, (40) tells us that under V 0 , player 1 is better off choosing x rather than e. At the
same time, (44) tells us that the rest of the players are also better off when player 1 chooses x
instead of e. Thus a vector-inferior collection of efforts (x, e, . . . , e) instead of (e, . . . , e) (which
in turn is lower than the first best (e∗ , . . . , e∗ )) leads to a Pareto-improvement. Under our
assumptions, this cannot be.
To see this more formally, note that if x < e < e∗ , then
(9.69) U1 (e, V 0 ) = U1 (e, V) > Ψ(ū, x, e)
as long as Ψ(ū, x, e) ≤ ū. By putting ū = U(x, V 0 ), Ψ(ū, x, e) = U1 (x, V 0 ), and using (44), we
contradict (45). This completes the proof of Claim [I].
Denote the maximal effort level among all symmetric equilibria under V 0 by ē. We have
already shown that ē < e∗ . We claim that
[II] max B(e; V 0 ) < e for all e ∈ (ē, L].
To see this, first observe that max B(L, V 0 ) ≤ L simply by definition. Now suppose that the
claim is false. Then for some e ∈ (ē, L], max B(e; V 0 ) ≥ e. Moreover, B(., V 0 ) is convex-valued
and upperhemicontinuous. But this establishes (using a simple argument analogous to the
Intermediate Value Theorem) the existence of some e0 > ē such that e0 ∈ B(e0 ; V 0 ). Moreover,
a fixed point of B(, ; V 0 ) corresponds to a symmetric equilibrium. But this contradicts the
definition of ē as the largest symmetric equilibrium effort level under V 0 .
To complete the proof of the proposition, suppose that there exists e ∈ (ē, L] which belongs to
B(e; V). Then max B(e; V 0 ) ≥ e > max B(e; V 0 ), where the first inequality follows from Claim
[I], and the second inequality follows from Claim [II]. This is a contradiction, and the proof
is complete.
CHAPTER 10
Notes on Inequality and Growth
Begin with the conceptual links between inequality and growth. I will say a little bit on each.
10.1 Inequality, Capital markets and Growth
We’ve already seen some of this from the earlier lectures. If capital markets are imperfect,
inefficiencies will arise from the fact that some individuals will be unable or unwilling to
take the socially optimal action — the imperfections generate a high or infinite utility cost
of doing so. Whether or not the final outcome is truly inefficient in the sense of dynamic
Pareto-efficiency remains to be seen (Mookherjee and Ray (2003) present some results to this
effect) but it is certainly the case that output is lower than in the first-best.
Just what is needed to translate this sort of result into a lower rate of growth is unclear. Ideally,
one would have to write down a model that allows for unbounded expansion — say, at some
constant exponential rate — and then examine whether such an expansion rate is affected
by the extent of capital market imperfections.
Such a model will also have to take account of the fixed cost barrier to setting up in a
new occupation or business. This must also “move” with growth, otherwise all credit
market imperfections must ultimately die away. One reason why such barriers may move
in step with growth is that barriers are denominated in human capital (e.g., education for an
occupation or professional staff for a business) which is getting more expensive with growth.
for a related model, see Rigolini (2003).
10.2 Inequality, Public Allocation and Growth
With high inequality, high-income or high-wealth individuals may garner a disproportionate

share of public assets. But doing so may be inefficient from the point of view of future growth.
Here are three examples:
1. Bhagwati and Desai (1956), in their study of Indian planning, suggest that the richer
business interests could easily buy up licences in a corrupt system, leading to the allocation
208 Notes on Inequality and Growth
of valuable resources (such as import quotas) in their direction. New blood didn’t get a
chance. See also Banerjee (1997) on bureaucratic corruption.
2. Sokoloff and Engerman (2000) and Acemoglu, Johnson and Robinson (2001) have argued
that an inefficient elite, once installed in power, might attempt to do all it can to keep the
non-elite at bay. As Acemoglu (2005) argues, they might set fiscal policy for three reasons: (i)
to generate revenue off the back of non-elite businesses, (ii) to make it more difficult for such
non-elites to conduct business and this way lower the factor costs of production (e.g., wage
rates) for their own use, and (iii) they might tax non-elites in order to reduce the resources
available for political opposition. This is more a theory of how inequality in the access to
political power (which may in turn stem from economic inequality in the past) might stifle
overall growth.
3. Esteban and Ray (2005) argue that public allocations may be severaly distorted even in the
absence of any corrupt motive. Imagine a world in which governments — just like private
agents — are unsure which direction the economy will, or should take. for instance, they
may not know whether textiles or software or call centers will be the wave of the future. If
the government has a limited supply of “licences” (infrastructual support, subsisdies, tax
breaks, etc.) that it can hand out, then it is rational for them to do so to sectors that lobby
hard. But hard lobbiers can be the outcome of individuals being highly productive (in which
case the government should support the lobby) or people just being rich (in which case there
is nothing to be gained from support). Thus high inequality distorts the lobbying signal,
and reduces the profitability of public allocation.
10.3 Inequality, Redistribution and Growth
The models of this section are diametrically opposed to the models of the previous section. In
the previous argument, high inequality leads to greater amount of rent-seeking, exploitation,
or regressive taxation of the non-elite, which in turn stifles growth through bad allocation.
In contrast, the models of Alesina and Rodrik (1994) and Persson and Tabellini (1994) argue
that high inequality sets up a clamor for redistributive taxation. That redistribution then
ends up being bad for growth because distortionary taxes stifle investment and therefore
growth.
These models are set up in a democratic context, and the median voter theorem is used in
particular. The simplest way to see these models (here I follow Bénabou (2000)) is to imagine
that income is distributed lognormally — ln y ∼ N(m, ∆2 ). Then it is trivial to see that the
log of the median income is precisely m, and also, if µ denotes mean income,
∆2
(10.1) m = ln µ − .
2
Suppose that there are only two policies on the table: t = 0 (laissez faire) and t = 1 (with full
lumpsum transfers to everyone; full redistribution). Everyone below the mean income will
prefer full redistribution, so that the degree of support for full redistribution may be proxied
by the value of the cfd evaluated at µ. Recalling (10.1) and using Ψ for the cdf of a standard
Notes on Inequality and Growth 209
normal, we see that the degree of support for redistribution is proportional to

∆2 /2 ∆
!
p=ψ =ψ .
∆ 2
Of course this value is bigger than 1/2, but that isn’t necessarily enough for full redistribution
to be passed. After all, as Bénabou argues, a large fraction of the individuals at the lower end
may not vote at all. For instance, if the π poorest agents don’t vote, then majority requires
that we need a support of (1 − π)/2 counting from the bottom of those individuals that do
vote. Adding back the π, this shows that the identity of the modified median voter is given
by λ, where Ψ(λ) = (1 + π)/2.
However, what appears (seemingly to be) robust is the point that ∂P/∂∆ > 0. That is, the
greater is the amount of inequality in the system, the more widespread is the support for
full redistribution.
What if one augments the model — as one needs to, to connect it with growth — by
presuming that redistributions are costly? Continuing to follow Bénabou, the simplest
depiction of that cost is one in which every income y is reduced equiproportionately from
y to ye−B , where B > 0, whenever redistribution takes place. Now the threshold income Y
which will support full redistribution is no longer µ, but is given by
∆2
!
ln Y = ln µ − B = m + −B ,
2
using (10.1). Consequently, the new degree of support is now given by
−B + ∆2 /2 B ∆
!
p=ψ =ψ − + .
∆ ∆ 2
Of course, this reduces the level of support for redistribution (relative to the benchmark
case) but what is more relevant is that the sensitivity of demand for redistribution to the
underlying amount of inequality becomes even higher. To see this, observe that
∂P B ∆ B 1

=ψ − +
0
+
∂∆ ∆ 2 ∆2 2
and this exceeds the ealier sensitivity, which is just
1 0 ∆

ψ
2 2
on two counts, provided that Y > m.1
10.4 Inequality, Lack of Redistribution and Growth
Notice that the model of the previous section relies on very different predictions regarding
the effect of redistribution on growth, relative to the arguments that went before. The
Alesina-Rodrik view is that redistribution occurs via distortionary taxation and is therefore
necessarily growth-reducing. But our earlier models argue that redistribution is actually a
1This will be the case anyway is at least a majority of the population is required to achieve redistribution.
good thing, either because it prevents inefficient public allocation, or leads to lower rent-
seeking, or permits individuals to find their true economic niche even if capital markets are
imperfect.
This forms the basis of Bénabou’s argument, which is simply to entertain the opposite
presumption that redistribution is beneficial to growth. Let’s try out this idea on the very
simple exercise presented above. Then instead of y being reduced by redistribution, let us
suppose that it is raised to yeB , where B > 0. Now the threshold income Y which will support
full redistribution is given by
∆2
!
ln Y = ln µ + B = m + +B ,
2
oncea again invoking (10.1). Therefore, the degree of support is given by
B + ∆2 /2 B ∆
!
p=ψ =ψ + .
∆ ∆ 2
But now this affects the demand for redistribution, or rather, the way it reacts to inequality.
To see this, note that
∂P B ∆ B 1

= ψ0 + − 2+ .
∂∆ ∆ 2 ∆ 2
Now notice that for small ∆, the change in the support for redistribution (as inequality
increases) is actually negative. The reason is not hard to see. By our assumed independence
of B from ∆, we see that for very small degrees of inequality practically everyone benefits
from redistribution. So there is near-universal support for it. As soon as ∆ climbs, some
of this support goes away as people prefer to retain their own incomes with laissez faire
rather than succumb to full redistribution and undergo an increase in income which will not
compensate them for the original loss. [For very high levels of inequality the relationship
will again turn positive, because most people are now poor and the redistribution can only
benefit them all.]
The interesting thing about this model is that it yields the same inequality-growth tradeoff,
but for exactly the opposite set of reasons from the Alesina-Rodrik-Persson-Tabellini exercise!
In particular, assuming inequality is not too high, an increase in it actually lowers the demand
for redistribution, but lower redistribution is presumed to be bad for growth, so that the
same reduced-form relationship between inequality and growth is obtained.
10.5 Inequality, Conflict and Growth
This is a connection emphasized by Alesina and Perotti (1994), Benhabib and Rustichini
(1996), Svensson (1994) and others. According to this view, high inequality causes political
instability, protests, violent demonstrations, coups, and rioting. These factors in turn make
for low growth.
The literature on this issue is largely empirical (see, e.g., Perotti (1996)). The conceptual
distinctions between inequality and polarization drawn by Esteban and Ray (1994, 1999)
have yet to be usefully applied to instability, though see the recent work of Montalvo and
Reynal-Querol (2005) on ethnic polarization and civil war.
10.6 Inequality, Status and Growth
Suppose that individuals accumulate in part because they derive utility from overtaking
other individuals in the wealth distribution. Then very flat wealth distributions will imply
low densities generally, so that the status effects of investment are likely to be relatively
small at the margin. Conversely, if the distribution is very egalitarian, then small changes in
investment will have large status effects, leading to a higher rate of growth.
This is only half the story, though, because the collective investment decisions made by
everybody will determine the distribution of wealth. In a fully specified equilibrium, the
distribution of wealth and the rate of growth would be jointly determined. the iintuitive
argument thus suggests that there are multiple steady states, but across those states the rate
of growth and the extent of inequality are negatively correlated.
Here is a very simple model that exhibits this effect. Suppose that individual utilities are
given by
(y − x)1−σ
A + F̂(x(1 + r)),
1−σ
where y is initial wealth, x is investment (so y − x is consumption), r is the rate of return to
investment, and F̂ is the cdf of wealth in the “next generation”.
We study steady states, so suppose that everybody’s wealth grows at some common rate
g (though with different level coefficients). Define, then, a cdf F on the normalized variable
y(t)/(1 + g)t . So we may rewrite individual utility above as
(y − x)1−σ x(1 + r)
!
A +F ,
1−σ 1+g
where tomrrow’s wealth is now in normalized form. For this to be a steady state, the above
expression must be precisely maximized at x(1 + r) = (1 + g)y, or x = (1 + g)y/(1 + r). Writing
down the first-order conditions and making this substitution, we see that
1+r 0
−σ
−σ r − g

Ay = F (y),
1+r 1+g
and this solves out for F over various supports and for various values of g. It is easy to see
that the negative correlation between inequality and growth is borne out in this model.
10.7 Positive Connections Between Inequality and Growth
All the models above suggest that there should be a negative reduced-form relationship
between inequality and growth, though the channels of interaction are very different.
Are there any reasons for believing that inequality might be growth-enhancing in some
situations? There certainly are.
10.7.1 Setup Costs and Extreme Poverty. Imagine a very poor society in which everyone
has low income, too low to incur the setup cost of a potentially profitable business. Then
that business activity will be thoroughly undersupplied. Instead, if wealth is distributed
unequally, some of the the individuals will be able to start the business activity, which
may then have high-growth implications for society as a whole. To be sure, the additional
inequality thus generated may be intolerable, but that is not the issue here. Inequality will
be positively associated with growth.
10.7.2 Uneven Technical Progress, Inequality and Growth. The sources of technical
progress are inherently uneven. Some sector (such as software design, biotechnology, or
financial services) takes off, and there is a frenetic increase in demand for individuals with
these skills. The economy as a whole registers growth, of course. But this growth is highly
concentrated in a relatively small number of sectors. Then, at least over the short to medium
run, inequality will be positively associated with growth. Whether this effect persists in the
longer run will depend on the frequency of such technological “shocks” and the speed at
which subsequent intersectoral adjustments occur.
10.7.3 The Tunnel Effect. The above argument is related somewhat to the “tunnel” effect
described by Hirschman and Rothschild (1973). You are driving through a two-lane tunnel,
where both lanes are in the same direction and, guess what, you get caught in a serious
traffic jam. No car is moving in either lane as far as you can see. You are in the left lane and
your spirits are not exactly high. After a while, however, the cars in the right lane begin to
move. Do you feel better or worse? It depends on how long the right lane has been moving.
At least initially, you might feel that the jam has cleared further ahead and that your turn to
move will come soon.
It has been the experience of several developing economies that the level of inequality in the
distribution of income increases over the initial phases of development. The responses to
such a rise in inequality have been varied, both across economies as well as within the same
economy at different points in time, and they have ranged from an enthusiastic acceptance
of the growth process that accompanied the rise in inequality to violent protests against it in
the form of social and political upheaval. Such differences in the tolerance for inequality may
be explained with the help of the tunnel analogy. The individual’s response to an uneven
improvement will depend on his beliefs as to what it implies for his own prospects. If he
believes such a rise in others’ fortunes is indicative of brighter prospects for himself in the
foreseeable future, then he may make complementary investments, further enhancing the
growth correlations described in the previous section.
Of course, if such an improvement in the welfare of others were to persist for a sustained
period of time, without any improvement in one’s own welfare, initial acceptance of the
improved condition of others would soon give way to anger and frustration, as in the
tunnel example. Moreover, increased inequality may not be tolerated at all if the perceived
link between the growing fortunes of others and the individual’s own welfare is weak or
nonexistent. The greater the extent of segregation in society to begin with, the higher the
possibility of this outcome.
10.7.4 Inequality, Savings and Growth. A last positive connection between inequality
and growth may arise from the effect of inequality on the ability to save, and subsequently
on growth rates. the argument is based on a nonconvexity similar to the setup costs story
of a previous subsection, but the nonconvexity this time arises from minimum consumption
needs. This may cause the savings function to become convex, at least over some range, with
the corollary that higher inequality (in this range of incomes) may be associated with greater
savings rates. The higher savings rates would presumably translate into higher growth rates
via standard argument.
10.7.5 Symmetric HoldUps, Inequality and Growth. the last argument that I present
here, due to Banerjee and Duflo (2003), is different from all the others in that it argues
for a nonmonotonic relationship between inequality and growth. This is analogous to
the assertion of a Kuznets curve, though that curve is concerned with a nonmonotonic
relationship between inequality and per-capita GDP.
Return to the Alesina-Rodrik view, discussed in an earlier section. This view may be
thought of as a situation in which the poor essentially hold up the rich, demanding transfers
in exchange for participating in the growth process where the benefits are flowing directly
to the relatively rich (and so need to be transferred via redistribution). In the Alesina-Rodrik
view, these transfers create distortionary incentives at the margin, thereby slowing down
investment and growth.
Banerjee and Duflo attempt to argue that this sort of holdup may well be symmetric. Just
as the poor may hold up the rich, the rich may hold up the poor. Excessive equality may
create a situation in which the traditionally rich demand compensation for pro-poor growth,
thus degrading that growth via analogous distortionary effects — in just the same way as
redistributive taxation did. This sort of argument suggests an inverse association between
growth and changes in inequality in any direction (the compensating transfers). Banerjee and
Duflo argue that such changes are more likely when inequality is either very high or very
low, and passing to the reduced form, this might suggest an inverse-U relationship between
inequality and growth. The warning: don’t impose linear relationhips in the empirics.
10.8 Empirical Results
The arguments above all rely on different “structural assertions”: inequality clearly works
through a variety of channels on its way to a final impact on growth. An initial exercise,
then, is to simply address the reduced-form question: is inequality positively or negatively
associated with growth? or perhaps more boldly, do initial inequalities retard or encourage
subsequent growth?
What is a good proxy for “initial inequality”? We would like to get a handle on inequalities
in wealth or assets at the beginning of the time period, but data on these are notoriously hard
to come by. One proxy for wealth is inequality of income at that time, but we must recognize
that this is an imperfect proxy. Wealth inequalities at some date are, in a sense, the sum
total of all income inequalities up to that date, and there is no reason why the last of these
inequalities should adequately mirror the history of all its predecessors.
Another proxy for wealth inequality is the inequality in some (relatively) easy-to-observe
asset, such as land. Data on land inequality are easier to obtain, although they are plagued
with problems of their own. Of these problems, the most serious is the distortion created in
countries that are subject to a land reform measure through the imposition of land ceilings.
In such countries land belonging to a single individual or household may be held under a
variety of names, thus creating the illusion of lower inequality than there actually is. Aside
from this problem, land inequality can only be a good proxy for overall inequality in wealth
if agriculture is either significantly important in the economy (for the beginning of the time
period under consideration), or at the very least, has been of significant importance in the
recent past. Fortunately for our purposes, this is a condition that is adequately satisfied by
developing countries.
Alesina and Rodrik (1994) regress per capita income growth over the period 1960–85 on a
variety of independent variables, such as initial per capita income and a measure of initial
human capital. [these are well-known controls: the initial income variable is a proxy for
possible convergence effects, while the human capital variable is used as a proxy for the
endogenous growth effects studies by Uzawa, Lucas and others (see, e.g., Barro (1991)).
Indeed, as far as these variables are considered, Alesina and Rodrik use the same data as
Barro, which makes for cleaner comparison with existing work. In addition, they included
data on initial inequality of income and initial inequality of land.2
Their results indicate a substantial negative relationship between initial inequality and
subsequent growth. Particularly strong is the influence of the Gini coefficient that represents
the initial inequality in land holdings. Their results suggested that an increase in the land
Gini coefficient by 1 standard deviation (which is only an increase of 0.16 in this case) would
decrease subsequent economic growth by as much as 0.8 percentage points per year. [The
Gini coefficient on initial income is only significant at the 10% level.]
These results are unaltered once we allow for structural differences across democratic and
nondemocratic political systems. What is more, the democracy dummy is insignificant both
by itself and when interacted with the Gini coefficient on land. It does appear that political
systems play little role in this relationship, so it is unclear whether a median-voter type
argument is central here (more on that below).
The Alesina-Rodrik findings are confirmed with the use of a more comprehensive data set
in Deininger and Squire (1996). Initial land inequality is more significant than initial income
inequality and stays that way even under several variations on the basic regression exercise
(such as the use of regional dummies). The insignificance of the political system also holds
up under the Deininger-Squire investigation.
The study by Persson and Tabellini (1994) finds slightly different results with somewhat
different specifications. They use the income share of the top quintile as their measure of
inequality. They don’t use land distributions. They carry out two kinds of regressions.
One is a set of historical regressions for nine now-developed countries for which the data
stretches well back.3 The other is a cross-section study, where the independent variable is
the share of the middle class (see also Perotti (1996)).
. . . to be completed
2The income inequality data are taken from Jain (1975) and Fields (1989). The land distribution data are drawn
from Taylor and Hudson (1972).
3The countries are Austria, Denmark, Finland, Germany, the Netherlands, Norway, Sweden, the UK and the US.
CHAPTER 11
Notes on Credit Markets
11.1 Introduction
We now turn to a detailed study of credit markets. As explained in my text, most developing
countries have a preponderance of informal credit markets (where the lenders are not formal
financial institutions). We will take the following route.
[1] We begin with a simple study of interest rate variation. We show how — depending on
the presence or absence of collateral — interest rates can be either high or low. Additional
fuel for this sort of variability will also be found when we study interlinkage.
[2] We then turn to a study of quantity restrictions on credit. These may actually take
two forms: credit rationing or loan pushing. The former involves the borrower wanting a
larger loan than what he is given at the going rate of interest. The latter involves wanting a
smaller loan. Finally, note that credit rationing can take the “micro form” that we have just
described or can be of a more “macro variety”: some people are entirely excluded from the
credit market. We will discuss various sources of credit restrictions.
[3] We study the general equilibrium of credit markets. In the analysis so far, we take as
given an outside option for borrowers. To be sure, this outside option will be endogenous in
a larger context. We show that the presence of information regarding past dealings is crucial
in determining these outside options, and indeed, that in some cases the credit market may
entirely collapse.
[4] We study interlinkage and segmentation in credit markets, using largely rural examples.
The idea is to interlink contracts to get around limited collateral, or limited incentives, by
offering a package. We will see that there may be several advantages to doing so. First,
some distortions can be avoided. Second, incentives (such as the threat of termination)
can be made to do double-duty (for, say, credit repayment as well as tenancy). Finally,
in the absence of monetizable collateral (such as unskilled labor or a small piece of land),
interlockers may have a serious edge over non-interlinking moneylenders.
216 Notes on Credit Markets
[5] Finally, we study group lending schemes such as those promoted by the Grameen Bank.
We show that under some conditions, group lending can be used as an effective tool ton
transfer information from borrowers to lenders, thereby making some new contracts feasible
(and profitable) for both lender and borrower.
11.2 A Simple Example
Imagine that loans are forthcoming at an interest rate of 10%, and that there are alternative
projects, each requiring a startup cost of 100,000 pesos. Suppose that the projects are arrayed
in terms of their rate of return, and that there are two of them, with rates of return pegged
at 15% and 20%. If there is no uncertainty about the projects, and all projects pay off fully in
the next time period, this is tantamount to saying that the projects will return gross revenues
of 115,000, and 120,000 pesos respectively.
Observe that in this case there is a perfect coincidence of interests between the bank and the
borrower. The bank wants its 10% back, and would presumably also want the borrower to
take up the optimal project. Given that the borrower wants to make as much money as he
can, there is no reason for him not take up the project with a 20% return. Everyone is happy.
But now let us change matters around a little bit. Suppose that the return to the second
project is uncertain. Thus keep the second project just the same as before, but suppose that
the first project pays off 230,000 pesos with probability 1/2, and nothing with probability
1/2. The expected return is just the same as it was before.
Now let us think about the rankings of these projects from the viewpoints of borrower and
lender. To do this, we shall assume limited liability: if a project fails, the borrower cannot
return any money to the lender: she simply declares bankruptcy. The bank would like to
fund the 20% project, just as before; indeed, more than it did before. This is because the 20%
project pays off its interest (and principal) for sure, while this happens only with probability
1/2 in the case of the 15% project.
What about the borrower’s expected return? Assuming that she is risk-neutral like the bank
is,1 it is 120, 000 − 110, 000 = 10, 000 for the safe project, and it is (1/2)[230, 000 − 110, 000] +
(1/2)0 = 60, 000 for the risky project. Her expected return is much higher under a riskier
project with a lower rate of return! She will therefore try to divert the loan to this project,
and this will make the bank very unhappy.
What went wrong with the market here? What is wrong is that the borrower has limited
liability. In this example, she pays up if all goes well, but if the project fails, he does not
repay anything (she does not get anything either but that is not the main point). In a sense,
this creates an artificial tendency for a borrower to take on too much risk: she benefits from
the project if it goes well, but is cushioned on the downside. The bank would like to prevent
this risk from being taken. Often it cannot.
We will come back to this scenario in more detail later. It will form the basis for theories of
credit rationing.
1A similar example can also be constructed even if the borrower is risk-averse.

Notes on Credit Markets 217
Observe that if the borrower could somehow be made to repay the loan under every
contingency, we would be back to a world that’s equivalent to one of perfect certainty.
The bank would not care what the borrower did with the money, and the borrower would
choose the project with the highest expected rate of return. But who can repay in all (or
most) contingencies? They are the relatively rich borrowers, who can dig into their pockets
to repay even if the project goes badly. We see here, then, in particularly stark form, one
important reason why banks discriminate against poor borrowers.
Thus institutional credit agencies often insist on collateral before advancing a loan. For a
bank which is interested in making money, this is certainly a reasonable thing to do. For
poor peasants, however, this usually makes formal credit an infeasible option. It is not that
they lack collateral to put up. But the collateral is often of a very specific kind. A farmer may
have a small quantity of land that he is willing to mortgage. But a bank may not find this
acceptable collateral, simply because the costs of selling the land in the event of a default is
too high for the bank. Likewise, a landless laborer may be seeking funds to cover a sudden
illness in the family, and pledge his labor as collateral: he would work off the loan. But no
bank will accept labor as collateral.
11.3 Interest Rate Variations
Classical explanation for a high rate of interest in informal credit markets rests on the
lender’s risk hypothesis (Bottomley [1963]). because there is risk, as in the previous section,
and because lending costs need to be covered, the lender tacks on a premium over and above
the opportunity cost of lending. This is a fairly obvious point but it should also be noted that
the point is often wrong! The reason is that the interest rate premium will also systematically
affect borrower behavior. There may be effects working through adverse selection, moral
hazard, or the strategic repayment incentive. In the next section, we begin our study of these
effects with the model of Stiglitz and Weiss (1981).
11.4 Quantity Restrictions
11.4.1 Quantity Restrictions Based on Adverse Selection. We base this on Stiglitz and
Weiss [1981]. Assume away the strategic default problem. But there is involuntary default
coupled with limited liability. To model this, assume that borrowers differ in some parameter
θ measuring riskiness. Project returns R are given by a cdf F(R, θ), and it is assumed that
higher values of θ generate the same means but higher risk. That is, if θ > θ0 , then
Z ∞ Z ∞
(11.1) RdF(R, θ) = RdF(R, θ0 ),
0 0
while
Z y Z y
(11.2) F(R, θ)dR ≥ F(R, θ0 )dR
0 0
for all y with strict inequality holding for some y.

Assume that individuals know their own type, while the bank only knows the distribution
of types and cannot recognize the individual riskiness of persons.
The project has a startup cost of B. There is limited liability: each individual has access
to only some fixed collateral C. Now study the repayment decision. If the rate of interest
charged is r, then an individual repays if and only if
(11.3) R − (1 + r)B ≥ −C,
where we are assuming that the bank can seize both the total return R as well as the collateral.
It follows that the borrower’s return (when the realized return is R and the rate of interest is
r) is given by
(11.4) π(R, r) ≡ max{R − (1 + r)B, −C},
which is convex in R, while the lender’s return is
(11.5) ρ(R, r) ≡ min{R + C, B(1 + r)},
which is concave in R.
This permits us to display the expected payoff to a borrower of type θ; it is
Z ∞
(11.6) π̂(θ, r) ≡ π(R, r)dF(R, θ),
0
while the expected payoff to the lender (when facing a borrower of type θ) is
Z ∞
(11.7) ρ̂(θ, r) ≡ ρ(R, r)dF(R, θ).
0
Now here is the important point. To be sure, the expected payoff to a lender — ρ̂(θ, r) — is
surely decreasing in θ, which is immediately intuitive. But a bit more subtle than this is the
observation that the expected payoff to a borrower — π̂(θ, r) — is actually increasing in his
own riskiness θ.
To put this another way, define, for each r the threshold θ̂(r) given by
(11.8) π̂(θ̂(r), r) = 0.
Then the set of types applying for a loan at the rate of interest r is precisely {θ : θ ≥ θ̂(r)},
and θ̂(r) is increasing in r. So increasing r lowers the average quality of loan applicants. It follows
that bank profits will — in general — be nonmonotonic with respect to r.
We may build on these concepts to define — relatively informally — a notion of competitive
equilibrium. To do this, let G stand for the distribution of types. First, define the profit rate
at r by
Z ∞ !
1
(11.9) d(r) ≡ ρ̂(θ, r)dG(θ) ,
B[1 − G(θ̂(r))] θ̂(r)
where ρ̂(θ, r) — which is expected profits when the borrower is of type θ — is defined in
(11.7), and θ̂(r) — the marginal borrower when the interest rate is equal to r — is defined by
(11.8).
r r
Dr
Dr
Sr Sr
S, D S, D
Sr
Dr
S, D
Figure 11.1. Credit Rationing: Three Cases
With free entry and exit of banks, the profit rate r will actually turn out to be the deposit rate
that is paid to depositors (which is why we are using the d-notation).
Now for each deposit rate d, there is a supply of deposits Ŝ(d) from potential depositors. Let
us assume that Ŝ(d) is an upward-sloping function. However, the supply of deposits viewed as
a function of r will typically be nonmonotonic simply because d is generally a nonmonotonic
function of r. Define S(r) ≡ Ŝ(d(r)); then this is the effective supply of deposits as a function
of the interest rate charged to borrowers (with the zero-profit condition already built in).
Finally, there is a demand curve for loans D(r), which has the usual downward-sloping
shape. Indeed, in our specific model, everyone who borrows borrows the same amount B,
while the cutoff θ̂(r) gets progressively higher with r. That means that the aggregate demand
has the appropriate shape in r.
Define a competitive equilibrium to be a compact set of interest rates charged by banks such
that no bank wants to deviate from those interest rates given optimal borrower behavior.
Proposition 11.1. Credit Rationing. (a) Suppose that there is an interest rate r with S(r) = D(r)
such that S(r0 ) < S(r) for all r0 < r. Then there is no equilibrium with credit rationing.
(b) Suppose that for every interest rate r such that D(r) = S(r), there exists r0 < r with S(r0 ) > S(r).
Then there is credit rationing in equilibrium.
Figure 11.1 illustrates. In panel A, there is no crossing of the two curves and so condition
(b) of the proposition holds trivially. In panel B, there is a crossing of the two curves but
condition (b) is satisfied nevertheless. In panel C, condition (a) holds. Thus in cases A and
B, there is credit rationing and in case C there is not.
It is easy to see how the proposition is proved. First assume case (a). Suppose, on the
contrary, that there is an equilibrium with credit rationing. Let r∗ be the lowest equilibrium
rate; certainly credit is rationed there. Observe that all r0 in the equilibrium set of interest
rates — call it E — must be equally attractive to depositors, so that
d(r0 ) = d(r∗ ) ≡ d∗ for all r0 ∈ E
and that the total supply of deposits divided across all lending banks must be S(r∗ ).
Fix r as given by condition (a).
First suppose that r∗ > r. Then we have S(r∗ ) < D(r∗ ) < D(r) = S(r), so we conclude that
S(r) > S(r∗ ) and consequently that d(r) > d(r∗ ). But now this equilibrium can be broken by
having a bank deviate to r from r∗ . Because d(r) > d(r∗ ), it can offer some d0 > d(r∗ ) and less
than d(r) and so make strictly positive profits. All the borrowers will come to him and all
the depositors as well. This contradicts the assumption that we are at an equilibrium.
Therefore r∗ < r. But then by (a), we know that S(r) > S(r∗ ), so d(r) > d∗ , the equilibrium
deposit rate. Now consider a deviation by a bank to charging an interest rate r and a deposit
rate d ∈ (d∗ , d(r)). All the depositors will come to this bank, so the supply of funds is not a
problem. To see if borrowers will be available, proceed as follows.
The total supply of loans to all banks is S(r∗ ) (depositors get the same rate in all banks
and simply divide themselves up). In particular: add up (or integrate) all depositor funds
available in equilibrium at interest rates r or less: this is not any more than S(r∗ ). On the
other hand, the total demand for loans just by risk types above θ̂(r) is D(r) = S(r) > S(r∗ ).
Thus demand (by these types alone) strictly exceeds total supply at interest rates r or less.
Indeed, because banks that ration credit randomly reject applicants, we see that conditional
on θ > θ̂(r), the distribution of risk types by borrowers is no different from G. These
borrowers will come to our deviating bank, and the bank can make positive profit from
them. We therefore have a contradiction, and the proof of part (a) is complete.
Now assume that the condition in (b) is true. If the conclusion is false, there is an equilibrium
with interest rate r, such that S(r) = D(r). We know that S(r0 ) > S(r) for some r0 < r. Because
S(r0 ) > S(r), we know that d(r0 ) > d(r). Therefore a bank can deviate by offering r0 and some
d0 > d(r) and less than d(r0 ) and still make profits. All the borrowers will come to him and
all the depositors as well. This contradicts the assumption that we are at an equilibrium.
It still remains to prove that an equilibrium exists with credit rationing in the case where (b)
holds. To this end, define r̄ as the maximizer of d(r) (there could be many but let’s assume
for ease of exposition that there is only one). Let r1 denote the lowest Walrasian outcome
(set it equal to infinity if there is no such outcome). If r̄ is smaller than r1 , it is easy to check
that all banks announcing r̄ is an equilibrium, and that it involves credit rationing.
The case r̄ = r1 is not consistent with case (b), so it remains to look at the case in which
r̄ > r1 . By condition (b), there is some conditional maximizer r∗ < r1 of d(r) (subject to
the constraint that r ≤ r1 ). Now define r2 to be the smallest value of r > r1 such that
d(r2 ) = d(r∗ ) (such an r2 must exist because r∗ is a local but not a global maximizer). Because
S(r2 ) = S(r∗ ) > S(r1 ) = D(r1 ) > D(r2 ), there must be excess supply at r2 . Now figure out a
division of banks among the two rates of interest so that all demand is soaked up. It is easy
to check that this forms an equilibrium. Notice that in this equilibrium there must be credit
rationing because there is a multiplicity of interest rates.
Some Observations.
(1) The absence of collateral, is, of course, critical to this sort of reasoning. But even if there
is some collateral (not necessarily enough to cover the entire risk of default), the collateral
can be used in a clever way to achieve some screening. Thus think now of a contract as a
pair (r, C), where r is the rate of interest and C is the collateral that has to be put up in order
to obtain loans at that rate of interest. Then the return to a borrower of type θ is
Z ∞
π̂(θ, r, C) = max{R − (1 + r)B, −C}dF(R, θ)
0
Z ∞
= −CF ((1 + r)B − C, θ) + [R − (1 + r)B]dF(R, θ).
(1+r)B−C
Using this expression, look at the tradeoff between r and C that leaves a borrower of type θ
on the same indifference curve. This is just
dr ∂π̂/∂C F ((1 + r)B − C, θ)
|π̂=π̄ = − = .
dC ∂π̂/∂r −B [1 − F ((1 + r)B − C, θ)]
Look at the case in which (1 + r)B − C is a “low” return; e.g., assume that (1 + r)B − C is smaller
dr
than the average value of R and that all distributions are symmetric. In this case, dC |π̂=π̄
goes up as θ goes up. That is, low-risk types are less willing to accept a given interest rate
increase for a given collateral reduction. This is because they fear less the loss of collateral,
and expect to pay interest in more states of nature.
It follows that if banks are offering (r, C) pairs, separation will be possible. This is the criticism
levied by Bester [AER 1985].
(2) Credit rationing is not inevitable in this model. For instance, if the supply of loanable
funds inccreases, credit rationing disappears. Alternatively, if projects were divisible and
there were constant returns to scale then we would get equal treatment: all applicants would
get a loan, though not as much as they would like — back to micro credit rationing.
(3) Notice that mean and riskiness are uncorrelated in this model. If the correlation is
positivem then the adverse selection effect would be weaker. Thus this form of credit
rationing relies on specific distributional assumptions.
(4) This form of credit rationing may also not be empirically very plausible if there are
several observationally distinguishable groups (along riskiness lines). Then at most one
group would be rationed (Riley [AER 1987]).
11.4.2 Quantity Restrictions Based on Moral Hazard.
11.4.2.1 Moral Hazard in Project Choice Let’s begin with a variant of the Stiglitz-Weiss
model in which there is unobserved project choice (and therefore moral hazard) on the part
of the borrower. Say that there are borrowers with various levels of (observable and seizable)
collateral C. There is also a whole range of projects indexed by θ as before, but this time no
particular θ is attached to any borrower; they all choose from this set of projects.
Assume that project returns are binary for every θ. Specifically, suppose that the return is
Q(θ) — an increasing function — with probability p(θ), and is 0 with probability 1−p(θ). The
gross expected value of project θ is, then, Q(θ)p(θ). We suppose without loss of generality
that p(θ) falls with θ; if this did not happen over some range, those projects would be never
be chosen anyway.
Assume, just as before, that each project requires the same loan size of B.
Now a borrower who puts down collateral C and faces a rate of interest r will choose θ to
maximize
p(θ) [Q(θ) − B(1 + r)] − 1 − p(θ) C,

where I assume that the best Q(θ) exceeds B(1+r), otherwise the borrower would not borrow
at all.
This determines θ(C, r) as a function of C and r; assume that the maximized value is unique
(it is easy to give suficient conditions for this to happen).
Observe that θ(C, r) is decreasing in C and increasing in r: more collateral induces safety; higher
interest rates induce risk.
To prove this, define Z = B(1 + r) − C, and convince yourself that the borrower maximizes
p(θ)[Q(θ) − Z]
by choosing θ. Let Z1 > Z2 and θ1 and θ2 be the corresponding maxima. Then by the
assumed uniqueness of the maximum,
p(θ1 )[Q(θ1 ) − Z1 ] > p(θ2 )[Q(θ2 ) − Z1 ],
while
p(θ2 )[Q(θ2 ) − Z2 ] > p(θ1 )[Q(θ1 ) − Z2 ].
Adding these two inequalities, we can conclude that
p(θ1 ) − p(θ2 ) (Z1 − Z2 ) < 0.

We can therefore conclude that p(θ) is increasing in Z, which completes the proof.
Now look at a person with C = 0. Notice that as long as
max p(θ(0, r))(1 + r) < 1 + r̄

r
where r̄ is the opportunity rate of return to funds, such people will be completely rationed
out, and all collaterals that are small enough will, by continuity, be completely rationed out
as well. There is no rate of interest at which a loan will be forthcoming to them.
So moral hazard also generally leads to credit rationing. Under some conditions, it is also
possible to predict what happens to the interest rate for different sizes of collateral. For
instance, under competitive lending, it must be the case that
p(θ(C, r))B(1 + r) + [1 − p(θ(C, r))]C = B[1 + r̄].
We need to first understand how a solution r is determined, for the left hand side of this
equation is nonmonotonic in the rate of interest. Begin at r = r̄; then surely the LHS is less
than the RHS. Send r to infinity; then everyone defaults for sure, so that p = 0 and once
again the LHS is smaller than the RHS. Whether or not there is some r such that the equality
holds is unclear (we just argued that it may not happen). But the point here is that if such
an r exists, there could be more than one solution. In that case Bertrand competition among
lenders must guarantee that the equilibrium rate of interest charged is the lowest among all
solutions to the equation.
In particular, in equilibrium, the LHS as a function of r must be locally increasing.
Raise C; then by our previous result p(θ) goes up. Obviously, B(1 + r) > C so in summary,
the left hand side of the equality above must go up. To restore equality, the interest rate r
must fall. This uses our particular selection of the equilibrium rate of interest as detailed
in the previous paragraph. Therefore under competition the interest rate on loans moves
inversely with collateral.
With monopolistic lending, matters are not so clear, which makes sense. A monopolistic
lender may well take advantage of a highly collateralized (and therefore safe) borrower to
push up the interest rate. I leave it to you to examine whether r must rise with C over some
stretch of collaterals.
11.4.2.2 Costly Effort and the Debt Overhang We now do a similar exercise with costly
effort, to understand the nature and extent of welfare loss when there is moral hazard.
Consider an indivisible project which requires funds of amount L to be viable. Output is
binary, taking values of either Q (good harvest) or 0 (crop failure). The probability of a good
harvest is an increasing, smooth function p(e), where e is the effort level of the agent who
oversees the project. Effort cost is given by e.
First, consider the problem of a self-financed agent. If investment takes place at all, the effort
level is chosen so as to
(11.10) max p(e)Q − e − L

e
Assume that there is a unique optimal choice e∗ ; it is described by the first-order condition
1
(11.11) p0 (e∗ ) =
Q
This is the efficient, or first-best level of effort.
Now consider a debt-financed agent. Let R = (1 + i)L denote total debt, where i is the interest
rate. Let w < L denote the value of the borrower’s transferable wealth that can be put up as
collateral. Now the effort choice of a borrower facing a total debt R is given by
(11.12) max p(e)(Q − R) − (1 − p(e))w − e
e
Denote the optimal choice by ê(R, w), defined by the following first-order condition:
1
(11.13) p0 (e) =
Q+w−R
Observe that ê(R, w) is decreasing in R and increasing in w.2 A higher debt burden reduces
the borrower’s payoff in the good state, but not in the bad state, dampening incentive to
apply effort. A bigger collateral, on the other hand, imposes a stiffer penalty in the event of
crop failure, thus stimulating the incentive to avoid such an outcome.
The lender’s profit is given by
(11.14) π = p(e)R + [1 − p(e)]w − L
We want to do some welfare economics, so we look for the Pareto frontier of possible
payoffs: hold the lender’s expected profit at any given level π, and maximize borrower
payoff, subject to incentive compatible choice of effort level. Implicit in our formulation is
the assumption that the opportunity cost of funds is zero, which is entirely innocuous and
can be generalized without any problem. In determining equilibrium choices, we will treat
π as given, and will later see the comparative static effects of increasing it. The special case
where π = 0 represents a perfectly competitive loan market with free entry. Since lenders
can always choose not to lend, it makes sense to restrict attention only to cases where π ≥ 0.
This last condition, together with (11.14) (and the fact that w < L) immediately implies that
R > w. Using this to compare (11.11) and (11.13), we conclude that ê < e∗ .
Proposition 11.2. As long as the borrower does not have enough wealth to guarantee the full value
of the loan, the effort choice will be less than first-best.
This is the debt overhang: an indebted borrower will always work less hard on his project
than one who is self-financed. Again, you can prove this by a revealed preference argument
without making any assumption on the curvature of p in e.
The variables determined in equilibrium are R and e. Equations (11.14) (the isoprofit curve)
and (11.13) (the incentive curve) jointly determine the outcome. It is easy to check that
the locus described by each is negatively sloped. If the borrower works harder, the risk of
2If the size of collateral is positive, or if the borrower has some outside option, there will be a participation
constraint in addition to the incentive compatibility condition described here. However, the participation
constraint only places a ceiling on the interest rate, and will be non-binding if the values of collateral and outside
option are low. Hence, we drop it from our analysis.
Isoprofit Curve
Incentive Curve
^ E
R
e^ e* e
Figure 11.2. Equilibrium Debt and Effort in the Credit Market.
default is reduced, and R must be lower to hold down the lender’s profit at the same level.
On the other hand, a reduced debt burden increases the incentive to work hard.
Notice also that as we move downward along the incentive curve, the borrower’s payoff is
increasing. Lower debt (R) increases borrower payoff for any given choice of effort, and hence
also after adjusting for optimal choice. If there are multiple intersections, only the lowest
among these (the one associated with the lowest R) is compatible with Pareto efficiency.
Further, the incentive curve should be steeper than the isoprofit line at the optimum point
(otherwise, a small decrease in R will increase both lender profit and borrower utility).
Figure 11.2 depicts a typical situation, point E representing the equilibrium.
We can now examine the comparative static effect of higher lender profit (π) or higher
wealth (w). Figure 2 shows the effect of increasing π. The isoprofit curve shifts up; in the
new resultant Pareto efficient equilibrium, the debt burden (R) increases, and so does the
interest rate (since the loan size is fixed), while the effort level falls.
Proposition 11.3. (Pareto efficient) equilibria in which lenders obtain higher profits involve higher
debt and interest rates, but lower levels of effort. Hence, these equilibria produce lower social surplus.
It is instructive to ask why higher rent extraction is associated with lower overall efficiency.
Lenders earn more profit by increasing the interest rate, which in itself is a pure transfer.
However, a greater debt burden reduces the borrower’s incentive to spend effort, increasing
the chance of crop failure and creating a deadweight loss. Consider two extreme cases. The
case of π = 0 represents perfect competition, and this situation generates the highest level
of effort among all. Notice, however, that since the debt burden still exceeds w, effort will
nevertheless be less than first-best. This tells us that the source of the inefficiency is not so
Isoprofit Curve
Incentive Curve
E'
^ E
R
^
e e* e
Figure 11.3. Effect of an Increase in the Lender’s Profit.
much monopolistic distortion created by the lender’s market power (although that certainly
exacerbates the problem), but the agency problem itself, and the distortion in incentives
created by limited liability. While the borrower shares in capital gains, he bears no part of
the capital losses (beyond the collateral posted). Working with other people’s money is not
the same as working with one’s own.
The other extreme case is that of monopoly. In this case, the value of π is maximized from
among all feasible and incentive compatible alternatives. In other words, the monopolistic
lender will choose the point on the incentive curve that attains the highest isoprofit curve.
The condition is the standard one of tangency between the two curves. This provides a
ceiling on the interest rate, or debt level (R̄), and the lender will not find it profitable to
raise it above this level. In more competitive conditions, this ceiling will still apply. If, in a
competitive credit market, there is excess demand for funds at R̄, the interest rate will not
rise to clear the market. We have an exact counterpart of Stiglitz-Weiss type of rationing
(rationing of borrowers, or macro-rationing in our terminology) in the presence of moral
hazard rather than adverse selection.3
The observation that borrower-friendly equilibria are more efficient has broad implications
for social policy. Any change which reduces interest rates, or improves the bargaining
power of the borrower will enhance effort and productivity. The latter involves institutional
changes, such as a reallocation of property rights over relevant productive assets from
lenders to borrowers, or an improvement in the latter’s outside options. Note, however,
3Stiglitz and Weiss (1981) discuss how their story can be recast as a moral hazard problem. However, the
incentive problem suggested there is somewhat different. With limited liability, borrowers will prefer a mean
preserving increase in spread in the distribution of returns. They will tend to pass up projects with secure
returns, and will instead select projects with high possible returns but also high risk. An increase in the interest
rate will reinforce this tendency.
that such policy interventions cannot result in improvements in Pareto efficiency — since
equilibrium contracts are by definition constrained Pareto-efficient — but result in higher
levels of social surplus. Yet, despite the fact that the gainers (borrowers) could potentially
compensate the losers, such compensations cannot actually be paid, owing to the wealth
constraints of the borrowers. Accordingly such policies will tend to be resisted by the losers,
and may not actually be adopted.
Can the model also generate micro-rationing—a situation in which even those who succeed
in obtaining credit still get too little? In other words, can there be under-investment in debt-
financed projects, in addition to under-supply of effort? We cannot address the issue in this
simple model, since the project has been assumed to be indivisible. However, it is easy to see
that the answer will be in the affirmative if the model is extended in a natural way to allow
for variable size of investment. Suppose output (when harvest is good) is Q(L), an increasing
concave function of the amount of loan or investment, but zero in the event of crop failure.
The complementarity between effort and investment will then generate suboptimal choices
on both fronts.4
In particular the phenomenon of nonlinear interest rates — where the interest rate depends
on loan size — may arise even when the credit market is competitive. An expansion in loan
size increases the debt burden, reducing the borrower’s stake in success, causing default risk
to increase. This may outweigh the effect of a larger scale of borrowing, making the lender
worse off. In order to remain commercially viable, the larger loan must be accompanied by
a different interest rate and/or level of collateral that reduces lender risk. Increases in the
interest rate can make matters worse, by raising debt burdens even further. While some loan
increases may thus be feasible if accompanied by higher interest rates, the lender may be
unwilling to lend beyond some level of loan size at any interest rate.5 Both micro and macro
forms of credit rationing can therefore arise, with credit ceilings depending on the collateral
that the borrower can post.
Long-term exclusive relationships and social networks can be useful in mitigating these
inefficiencies to some extent. When the lender and borrower enter a long-lived relationship,
it expands the opportunity for the lender to relax limits on the borrower’s current liability
by extracting repayment in future successful periods (by the institution of debt), or by the
threat of terminating the supply of credit (an issue further discussed in Dutta, Ray and
Sengupta (1989)). A similar reason underlies the role of lending within social networks,
where punishments can be imposed for loan defaults in other spheres of social interaction,
and third-party community-based sanctions can be brought to bear on defaulters owing to
the rapid flow of information within the community. More on this below.
11.4.2.3 A Consumption-Loan Variant of Debt Overhang Recall that a debt overhang

refers to a state of affairs in which an ongoing debt reduces the effort taken by the borrower
to ensure a good output. It is unclear whether one thinks of this as an instance of strategic
or involuntary default — both elements are present.
4It is easy to see that in any constrained efficient contract the loan size will be selected to maximize p(e)Q(L) − L.
5
For an explicit example, see Aghion and Bolton (1997).
Here is a variant of the debt overhang model that uses consumption smoothing. Let wealth
be w0 today, while tomorrow it has value w1 with probability p and 0 otherwise. Assume
that w1 > w0 . If L is borrowed today and then R repaid tomorrow (only in the good state, of
course), then expected utility is given by
(11.15) u(w0 + L) + δpu(w1 − R),
where we have normalized u(0) to 0.
The lender gets an expected value of
(11.16) −L + δpR,
where this formulation presumes that the risk-free rate of interest is given by δ = 1/(1 + r).
Now imagine that p is exogenous and that we are trying to find a Pareto-optimal allocation.
That is, maximize (11.16) while respecting some lower bound on the value in (11.15). It is
easy to see that we must set
(11.17) w1 − R = w0 + L,
so that there is perfect consumption smoothing (over those states that permit it). The exact
values of R and L cannot be pinned down until we know the particular Pareto-optimal point
that we are trying to characterize. For instance, if lenders operate in a competitive market
with free entry, then lender’s profits will be driven to zero, so that, by (11.16),
1
(11.18) R= L,
δp
which is just another way of saying that the rate of interest is driven down to the risk-adjusted
1
value δp − 1.
In this case, equations (11.17) and (11.18) fully pin down the first best credit scheme.
Now suppose that the probability of success can be influenced by the deliberate application
of effort. To keep things very simple, we shall suppose that there just two possible levels
of effort, 0 and 1. The corresponding probabilities of success are p(0) and p(1) respectively,
where p(0) < p(1). The cost of setting effort equal to zero is 0, while the cost of setting effort
equal to 1 is D.
Thus, ex post, imagine that there is an overhang of R and the borrower is deciding on effort
input. He will choose the high level of effort if
[p(1) − p(0)]u(w1 − R) ≥ D,
or equivalently, if
!
D
(11.19) R ≤ R̄ ≡ w1 − h ,
p(1) − p(0)
where h is just the inverse function of u.
Now if R̄ is higher than the first-best equilibrium, there is no problem. The first-best outcome
still obtains.
On the other hand, there may be credit rationing especially if wealth is low or if D is high.
In that case, one needs to compare the rationed contract with another possibility in which
e = 0 and loans are advanced at the higher competitive rate δp(0)
1
− 1. We would then expect
that no credit rationing would occur both among the very poor and the very wealthy, while
credit rationing would be observed for borrowers with intermediate levels of wealth.
We can easily extend these to cohesive notions of equilibrium (with free entry on the part of
lenders). Notice that such equilibria would be second-best Pareto optimal: the government
cannot do better if it must also repect the moral hazard constraint of the borrower.
However, this assumes that lenders can offer exclusive contracts. If exclusivity cannot be
enforced, then these equilibria will break down.
To see this, notice that at the credit-rationed equilibrium, a borrower is under rationed even
1
relative to the higher interest rate δp(0) − 1. The way to see this is to note that the loan repayment
amount with perfect insurance under the higher interest rate must exceed its counterpart
under the lower interest rate. Because this counterpart, in turn, exceeds the threshold R̄, it
follows that the credit-rationed borrower would love to take some additional loan (small
1
enough) under the higher interest rate δp(0) − 1 (or even at a rate a little higher than that).
If a lender supplies this loan — which he will if exclusivity cannot be monitored — then
this creates a negative externality on the “incumbent lender”, and second-best optimality (at
least relative to an ideal in which the government can enforce exclusivity) will break down.
11.4.3 Strategic Default. Consider the following model of working capital. Output simply
depends on the amount of loan (L) taken via the function F(L) satisfying standard properties.
A borrower is engaged in a repeated relationship with a lender, but if he leaves the
relationship he has an outside option given by a normalized value of v.
Consider only stationary contracts (L, R) for the moment (later, we explore the nonstationary
case). Then what is the incentive constraint for the borrower? It is given by the restriction:
(1 − δ)F(L) + δv ≤ F(L) − R,
or, rearranging:
(11.20) δF(L) − R ≥ δv.
Now suppose that given these outside options, the lender has all the bargaining power.
Suppose, more over, that the opportunity interest rate of funds lent out is r. Then he
maximizes his payoff
R − (1 + r)L
by choice of (L, R), and subject to the incentive constraint (11.20).
It is easy to see, either by constructing the Lagrangean and differentiating, or drawing a
diagram, that the solution involves the granting of a loan of size L̃, where this solves the
equation
(11.21) δF0 (L̃) = 1 + r,
and then asking for a repayment R̃ satisfying

(11.22) δF(L̃) − R̃ = δv.
Now let us ask ourselves: is there credit rationing or loan pushing at this solution? Notice,
to begin with, that L̃ is surely less than L∗ , which solves F0 (L∗ ) = 1 + r and which would
surely have been the optimal amount barring default considerations. So there is clearly a
restruiction of the loan relative to this value. But this does not say anything a priori about
whether there is credit rationing or loan pushing. To settle this we will have to look at the
implict rate of interest on the loan, which is r̃ ≡ R̃L̃ − 1.
The question is: would the borrower have preferred to borrow more or borrow less at this
rate of interest r̃?
The answer to this question depends on the discount factor of the borrower. First note that
a solution to the maximization problem exists if and only if
1+r
(11.23) max[F(L) − L] ≥ v.
L≥0 δ
Suppose that we are at (or very close to) the existence borderline. In that case,
R̃ ' (1 + r)L̃,
which is just another way of saying that
r̃ ' r.
At the same time, L̃ < L∗ . So obviously for low discount factors we have credit rationing
(relative to the implicit or explict rate of interest on the loan).
On the other hand, suppose that (11.23) holds for some threshold value of δ but that — in
any case — δ is very close to one. Then the structure of the optimal contract — see (11.21) —
tells us that
L̃ ' L∗ ,
while
R̃ > (1 + r)L̃.
[To understand this last inequality, recall from the optimality condition (11.22) that R̃ =
δF(L̃) − δv > (1 + r)L̃, where the strict inequality holds because (11.23) must hold strictly
when δ is close enough to one.]
In other words, for δ sufficiently close to unity there must be loan pushing.
It turns out that loan pushing is a property of situations in which the creditor has some
bargaining power and the borrower is sufficiently patient. For if we give the borrower all
the bargaining power in this model, then loan pushing can never occur. To see this, consider
the borrower’s problem when he has all the bargaining power: lender profits must be zero,
so the borrower maximizes
(11.24) F(L) − (1 + r)L
subject to his own incentive constraint (11.20), which we may rewrite here as
(11.25) δF(L) − (1 + r)L ≥ δv.
Again, consider two cases. First, suppose that

(11.26) δF(L∗ ) − (1 + r)L∗ ≥ δv.
In this case the borrower must choose the unrestricted maximum L∗ . Then there are no
quantity restrictions. Otherwise
(11.27) δF(L∗ ) − (1 + r)L∗ < δv.
But of course, it is still possible that (11.23) holds. On this assumption, the borrower will
choose the loan size L(δ), which is the maximum loan size L such that (11.25) holds. In this
case there will be credit rationing.
11.5 Information and Equilibrium in Credit Markets
The main point of this section is to argue, along with Ghosh and Ray [1996, 1998], that an
informational breakdown in informal credit markets is likely in the process of development,
and that this has implications for the structure of contracts, we well as for the very existence
of such markets.
Consider v, the outside option. Where does it come from? Presumably, this is the expected
(present-value) payoff conditional on the termination of an ongoing bilateral relationship.
The reason for this termination will typically matter. For instance, if a borrower has defaulted
and this is the cause of termination, then if this matter becomes known to new lenders they
may not be willing to advance a loan.6
Thus, v is some combination of punishment as well as the going equilibrium contract. There
are obviously many ways to model this. Let us take a particularly simple version, where
we assume that the lending market is competitive and borrowers have all the bargaining
power. To review the usual stuff, recall that given the outside option v and the opportunity
rate of interest r facing lenders, borrrowers maximize
(11.28) F(L) − (1 + r)L
subject to the incentive constraint
1+r
(11.29) F(L) − L ≥ v,
δ
where δ, it will be recalled, is the discount factor.
It is obvious, then, that in any solution L to the above problem, L must be at least as large as
the value L̂ that maximizes the LHS of the incentive constraint (11.29).
Let the optimum value of (11.28) be given by w (we emphasize that w depends on v).
Now let us try to endogenize v. Suppose that when a borrower defaults, he approaches a
lender every period thereafter. The lender checks on the borrower’s past and uncovers the
6Why new lenders may be unwilling to give a loan to past defaulters is actually a matter requiring some serious
study. For instance, can one sustain this sort of outcome without any incomplete information at all on the part
of the borrower? If there are a lot of borrowers and a lender can always find a fresh borrower, then this sort
of assumption is relatively easy to maintain. Otherwise the equilibrium supporting this sort of penalty may be
surprisingly convoluted — in the presence of complete information, at least.
default with probability p. In that case, the lender refuses the loan, and the borrower moves
to the next period where exactly the same story repeats itself. If, on the other hand, the
lender fails to uncover the default, the borrower enters into a new credit relationship with
the lender. But in that case, the present value of the contract is w.
It follows that v — the expected value of the outside option — is given by
1−p
(11.30) v = pδv + (1 − p)w = w.
1 − δp
Then we can write v = (1 − ρ)w, where
p(1 − δ)
(11.31) ρ≡
1 − δp
can be viewed as the scarring factor. Notice that if p gets very close to one, so that a default
is always recognized, then the scarring factor converges to one as well. On the other hand,
for any p strictly between zero and one, the scarring factor goes to zero as δ goes to unity.
Proposition 11.4. Define
(1 + r)(1 − δ)L̂/δ
(11.32) ρ∗ ≡ ,
F(L̂) − (1 + r)L̂
1+r
where L̂ is the maximizer of F(L) − δ L.
Then a competitive equilibrium exists if and only if ρ ≥ ρ∗ .
Thus a certain minimum degree of scarring is needed before a competitive equilibrium can
be guaranteed. This is intuitive. If scarring did not exist, individuals could always renege
on loans and dive back into an anonymous pool.
Notice by the way that ρ∗ ∈ (0, 1).
Proof of Proposition. Necessity. Suppose, on the contrary, that ρ < ρ∗ , but that a loan size L
constitutes a competitive equilibrium. Then
v = (1 − ρ)[F(L) − (1 + r)L]
= [F(L) − (1 + r)L] − ρ[F(L) − (1 + r)L]
> [F(L) − (1 + r)L] − ρ∗ [F(L) − (1 + r)L]
(1 + r)(1 − δ)/δ
= [F(L) − (1 + r)L] − [F(L) − (1 + r)L]
F(L̂)/L̂ − (1 + r)
(1 + r)(1 − δ)/δ
≥ [F(L) − (1 + r)L] − [F(L) − (1 + r)L]
F(L)/L − (1 + r)
(1 + r)(1 − δ)L/δ
= [F(L) − (1 + r)L] − [F(L) − (1 + r)L]
F(L) − (1 + r)L
= [F(L) − (1 + r)L] − (1 + r)(1 − δ)L/δ
1+r
= F(L) − L,
δ
where the first weak inequality in the string above uses the fact that L ≥ L̂. But now we have
a contradiction to the incentive constraint (11.29).
Sufficiency. Conversely, assume that ρ ≥ ρ∗ . Define ŵ ≡ F(L̂) − (1 + r)L̂.

For each w ≥ ŵ define v ≡ (1 − ρ)w. First, consider v̂ = (1 − ρ)ŵ. Notice that for v = v̂, the
maximization problem described by (11.28) and (11.29) has a feasible solution. For
1+r (1 + r)(1 − δ)
F(L̂) − L̂ = F(L̂) − (1 + r)L̂ − L̂
δ δ
= ŵ(1 − ρ )
∗
≥ ŵ(1 − ρ) = v̂.
Now, for each w ≥ ŵ, form the corresponding v (as we just did for ŵ) and keep noting the
borrower’s maximal w relative to that v. Call this mapping h(w). Notice that there exists
w̄ ≥ ŵ such that when we put v̄ = (1 − ρ)w̄,
1+r
F(L̂) − L̂ = v̄.
δ
Put another way, L̂ is the only solution satisfying the incentive constraint when the outside
option is v̄. That means the borrower’s maximum w = h(w̄) at this point is simply ŵ. So we
have found two points ŵ and w̄ such that ŵ ≤ w̄ and
h(ŵ) ≥ ŵ and h(w̄) = ŵ ≤ w̄.
Noticing that h is continuous is between, we have proved existence.
Thus is the scarring rate is too low, equilibrium falls apart. What might restore equilibrium
and still permit credit markets to function?
11.5.1 Credit Macro-Rationing. The first possibility is credit rationing at some macroeco-
nomic level. To see how this fits, suppose that a past defaulter may be excluded from future
loan dealings for two distinct reasons:
Targeted Exclusion. The defaulter’s nasty behavior is unveiled bya new lender (which
happens with probability p), and he is refused a loan. This is already present in the model
above and we add nothing here.
Anonymous Exclusion. Whether or not a potential borrower has actually defaulted in the
past, he may face difficulty in getting a loan. This is macro-rationing of credit, analogous
to the equilibrium unempoloyment rate in Shapiro and Stiglitz [1984]. Let us denote the
probability of such exclusion (in any period) by q. Notice that to build a coherent model in
which q > 0, we really have to answer the question of why the market is not clearing — why
lenders do not pounce upon available borrowers and ply them with loans. [Note, by the
way, that the same issues come up when we attempt to explain p > 0 without taking recourse
to any reputational factors.] One coherent model is given by the case in which lenders make
zero expected profits, so that they are always indifferent between lending and not lending.
The main point is that anonymous exclusion may be an equilibrium-restoring device. To

see this, let us calculate ρ, the effective scarring factor, when there is both targeted and
anonymous exclusion. The corresponding equation is
π(1 − δ)
(11.33) ρ≡
1 − δπ
where π, now, is the overall probability of being excluded at any date. It is easy to see that
(11.34) π = 1 − (1 − p)(1 − q).
Now notice that irrespective of the value of p, q can always adjust to guarantee that an
equilibrium exists. [To be sure, the determination of q becomes an interesting question, but
this is beyond the scope of the present exercise.]
11.5.2 Reputational Effects: Starting Small.
11.5.3 Nonstationary Contracts: Starting Small.

CHAPTER 12
Interlinked Contracts
We return to the question of borrower-lender matching. The data reveal that borrowers and
lenders are ususally engaged in “compatible” occupations: the landlord lends to his tenant,
the trader lends to farmers who supply to him, and so on. Moreover, one observes that the
contacts are ofen interlinked with other contracts. A loan is often offered as part of a tenancy
arrangement. A trader may advance loans to the farmers who supply to him. What explains
these interlinked contracts?
[1] Interlinkage as Assurance. Suppose that the lender’s principal occupation involves
increasing returns to scale. Trading is the best example, because the trader must incur large
fixed costs to transport the produce. Faced with uncertainty regarding the quantity of output
he can transport to the market during harvest time, he might find it in his interest to “tie up”
some farmers by giving them loans in exchange for the promise of output sales to him.
[2] Interlinkage as Enforcement. Interlinked relationships are sometimes useful for

preventing strategic default as well. To see this, it will be useful to recall two stories with very
similar features. First, think about the model of strategic default that we considered earlier.
We noticed that to prevent default, the moneylender cannot drive down the borrower to his
participation constraint. To avoid default, a certain surplus over the next best option had to
be provided. The borrower will trades off the loss in this surplus in future dates with the
one-time gain to be had from default.
Second, notice that a very similar story can be constructed for a tenancy contract (and we
will do that in the last section of this course). A tenancy contract can be supplemented
by intertemporal incentives which promise to renew the contract in the case of satisfactory
performance, and end it otherwise. This firing threat can only be a threat, however, if the
tenant earns by being fired than he does with his current (and long-term) landlord. Thus
the permanent contract must be in the nature of a carrot, which can be used as a stick in the
event of noncompliance with the contract.
The simple observation is that with an interlinked relationship, a single carrot can be used as
two sticks, as long as deviations cannot be carried out simultaneously on both fronts. Combine
the two scenarios. For instance, suppose that a landlord has a tenant to whom he offers a
rental contract with threat of eviction in case the output is lower than some pre-understood
236 Interlinked Contracts
minimum. Such a contract must, for the same reasons considered in the stories above, carry
with it a certain surplus. Now it is easy to see that a loan to the tenant can be supported by an
“interlinked threat”: if the loan is not repaid, then the tenancy will be removed. The surplus
in the tenancy thus serves a twin role. It assures the provision of appropriate effort in the
tenancy contract, while at the same time it doubles as an incentive to repay loans. In this
sense, the landlord is at a distinct advantage in advancing credit to his tenant, because he
has at his disposal a pre-existing instrument of repayment. In contrast, a pure moneylender
lending to the same tenant must offer additional incentives for repayment through the credit
contract itself.
[3] Interlinkage and Nonmarketable Collateral. We’ve already talked about the possible
nonmarketability of collateral such as labor or even land. This might create what looks
like interlinked contracts, because it will be very natural that someone who has use for the
borrower’s collateral will be lending to him. But one should be careful to note that this — in
itself — does not guarantee that the contract itself will be interlinked. But it does go some
way towards explaining market segmentation.
[4]Interlinkages and Distortions. The example that follows is based on Gangopadhyay

and Sengupta [OEP, 1987].
Suppose that a crop is grown with market price p. Farmers can produce the crop on their
own and market it through a trader, but with working capital they can produce an even
larger amount. One way to describe this is to assume that the famer has access to an indirect
production function F(L) defined on the amount borrowed, which is L. Think of F(0) as the
total output produced in the absence of working capital, assume that there are no other costs
of production, and define A ≡ pF(0). This is his outside option.
Now suppose that a trader has access to funds at some opportunity rate of interest i, and
can lend it to the farmer. In principle he can offer a contract of the form (q, r), where q is the
price at which he can pledge to buy output from the farmer and r is the rate of interest at
which he offers the loan.
If the trader offers a simple contract in which only some rate of interest r is specified, then
the farmer will choose L to maximize
F(L) − (1 + r)L,
yielding some return to the farmer — call it B. Note that B > A if, for instance, F0 (0) > 1 + r.
It will also yield a return C to the trader, where
C = (r − i)L(r),
(L(r) being the optimal loan taken by the farmer at the rate of interest r).
To compare the surplus generated — which is B+C — with the maximal total surplus that can
be generated, introduce a fictional business in which the trader and the farmer are actually
the same entity. Then this joint entity will choose L∗ to maximize
F(L) − (1 + i)L.
Interlinked Contracts 237
Let S be the maximal value. This is the total surplus available in the system. It is easy to see
that for any r > i, S > B + C. This is the distortion: the loan size L(r) is distorted away from
L∗ = L(i).
However, there is an interlinked contract which regains full efficiency, and the trader will
have the incentive to offer it. Define (q∗ , r∗ ) such that
(12.1) q∗ /p = (1 + r∗ )/(1 + i),
and such that
(12.2) max q∗ F(L) − (1 + r∗ )L = A.
L
It is easy to see that there is a unique contract satisfying these two properties. It must involve
q∗ < p and r∗ < i. That is, the loan is advanced at a rate below the opportunity cost of funds
for the trader, while the surplus from the loan is removed in the form of a lower output
price.
This contract must maximize trader profits. By (12.1), the farmer will choose a loan size
exactly equal to L∗ , so this is the undistorted amount. By (12.2), the farmer is given just
enough so that he will participate in the contract. So the trader will remove S − A, which is
the highest that he can possibly get.
It is in this sense that interlinkage can remove distortions.
While this sort of argument has insight, it needs to be refined. For instance, there are other,
noninterlinked contracts that are also efficient in this case. But these must be nonlinear
contracts. For instance, consider the forcing contract under which the trader says:
Either borrow L∗ at a high rate of interest R such that pF(L∗ ) − (1 + R)L∗ = A, or don’t borrow
anything at all.
This will also implement the first best. In other words, the interlinkage result may not be
robust to nonlinear contracts.
[5] The interlocker’s Edge? Of course, notice that this sort of forcing contract only makes
sense when there is no asymmetric information. The lender has to be able to calcuate the
first-best loan. However, as Ray and Sengupta [1989] argue, there are several situations in
which an interlocker has really no intrinsic advantage over a pure moneylender. To see this,
look at the following cases.
Case 1. The pure moneylender cannot offer arbitrary nonlinear contracts. In this case, as we
have seen in the discussion above, the interlocker has an advantage.
Case 2. The interlocker faces better terms in the market where he is active. In this case he
can do better than the pure moneylender, and this pure gain may be transformed into much
larger profits if there are epsilon costs of entry.
Consider the following example. Suppose that a laborer needs a loan to tide over the slack
season. In the peak season he may be employed with probability p or unemployed with
probability 1 − p. The moneylender cannot do anything about the laborer’s probability of
238 Interlinked Contracts
employment but a large landowner who will hire labor in the peak season certainly can.
This gives him the edge.
Case 3. Differential observability. Suppose a loan is given in the first period. In the second
period, there is a standard principal-agent problem: the output is produced and the loan
repaid. A fixed interest contract is not optimal then for the same reason that a fixed rent
contract is not generally optimal. Lender can do better if he can observe output and condition
repayments on these observations.
Case 4. Extend the trader-lender model so that the final output price is uncertain. If
the lender is risk-neutral and the tenant is risk-averse, then the trader can provide perfect
insurance by buying the crop at a fixed price. There is no need (for the trader) to observe the
final price of the output. But if the moneylender cannot observe this final output, he cannot
offer an equivalent contract.
Proposition 12.1. If Cases 1–4 do not hold, then interlinking and pure moneylending yield the same
returns.
Proof. See Ray and Sengupta [1989].

CHAPTER 13
Credit Policy
13.1 Group Lending
Begin with an adverse selection model along the lines considered by Stiglitz and Weiss, but
in even more simplified form (we base this on Ghatak [1999]). There are a continuum of
borrowers each seeking a unit loan. The repayment on the loan (principal plus interest) is
denoted by r. The opportunity cost of making the loan is denoted by i.
We consider only two possible outcomes: full repayment and no repayment. Borrowers are
indexed by their probabilities of repayment, which we denote by p, and the amount that
they repay conditional on success, which we denote by R. Without loss of generality we let
p index borrower type and we assume that pR(p) is a constant, R, for each type of borrower.
Borrowers are distributed with support on the interval [q, 1], where q > 0, and with density
g on this support.
Each borrower has an outside option of u, and everybody is assumed to be risk-neutral.
If we assume that
(13.1) R > i + u,
then it is first-best optimal to lend to all borrowers. But let us see how the equilibrium
works out. Assume that there is a competitive market among lenders, so that we impose
the zero-profit condition on them.
Suppose that the going rate of repayment is r. Then who will borrow? Only those types p
such that p ≤ p(r), where p(r) is given by the equation
R − p(r)r = u,
or
R−u
(13.2) p(r) = .
r
That is, the borrower pool will lie in the interval (q, p(r)]. It follows that the zero-profit
equilibrium condition for the lender is
(13.3) a(r)r − i = 0,
240 Credit Policy
where
R p(r)
q
sg(s)ds
(13.4) a(r) ≡ .
G(p(r))
where G is the cdf.
Proposition 13.1. If (13.1) holds, then there is r such that a(r)r − i = 0. Moreover, there is a unique
equilibrium (generically) given by the smallest of the solutions (in r) to this zero-profit equation.
Proof. Begin by looking at r = i. In that case, we see that even the type p = 1 will benefit by
taking a loan because of (13.1). It follows that p(i) = 1. But nevertheless, given that there is
a spread of borrowers, a(i) < 1. Therefore
(13.5) a(i)i − i < 0.
Now take r to the largest possible value which is consistent with the worst type borrower
borrowing; call this value ρ. Then
R p(r)
q
sg(s)ds
lim inf ra(r) = r lim inf .
r→ρ r→ρ G(p(r))
Notice that both numerator and denominator of this expression go to zero as r → ρ. So, using
L’Hospital’s Rule, we may conclude (taking derivatives in both numerator and denominator
above) that
pg(p)
lim inf ra(r) = ρ lim inf = ρq = R − u > i,
r→ρ p→q g(p)
using (13.1) again. Consequently

(13.6) lim inf a(r)r − i > 0.
r→ρ
Noting that a(r)r is continuous, we can conclude from (13.5) and (13.6) that there exists
r ∈ (i, ρ) such that a(r)r − i = 0. Notice that there is a compact set of such solutions, so that a
minimum — call it r∗ — is well-defined.
Notice that an equilibrium must belong to the set of zero-solutions. To see this, observe that
generically there exists an interval of the form (r∗ , R∗ +) such that a(r)r−i > 0 in this interval.1
Any bank can cut any other announcement to this value, get all the first pick at borrowers,
and make positive profits. So the equilibrium must involve a common announcement of r∗ .
Now we turn to the possibility of group lending. To begin with, simply look at lending
to groups of two (without worrying about whether this can be an equilibrium or not). A
contract is now a pair (r, c), such that r is the payment a borrower makes if he is successful
1Why is this generic? To see this, observe that the first zero of the function a(r)r − i must be a cut from “below”,
because a(i)i − i < 0. generically, it must be a cut in which case the condition in the main text is satisfied. There
is a nongeneric possibility that the first zero is a point of tangency between the a(r)r − i line and the horizontal
axis.
Credit Policy 241
and c is an extra payment that he makes if he is successful but his partner is not. So the
expected return to a type p when he is paired with a type p0 is
(13.7) π(p, p0 ) ≡ pp0 [R(p) − r] + p(1 − p0 )[R(p) − r − c] = [R − rp] − p(1 − p0 )c.
Of course, π2 (p, p0 ) > 0: everybody wants a safer partner. But the important point is that
π12 (p, p0 ) > 0 as well: the safer person benefits more from a safer partner:
π2 (p, p0 ) = pc,
π12 (p, p0 ) = c.
This means that in any group formation games, the matching of unqeual types can never be
a core outcome, because
(13.8) π(p, p) + π(p0 , p0 ) > π(p, p0 ) + π(p0 , p).
To see this, simply use (13.7) and grind away at the algebra — it’s easy. Thus group lending
gives rise to positive assortative matching.
Now we can study how joint liability contracts such as these may be used to overthrow
the individual liability equilibrium. Assuming — using the above argument — that similar
types will pair together, we may conclude that the return to a p-pair from a contract (r, c) is
(13.9) π∗ (p) ≡ [R − rp] − p(1 − p)c.
Now consider a variation of (r, c) around the single liability contract (r∗ , 0) such that r < r∗ ,
c > 0, and which leaves the marginal type (under the old equilibrium) — p(r∗ ) unaffected.
Writing p∗ = p(r∗ ), this means (using (13.9)) that
∆(p∗ ) ≡ −p∗ dr − p∗ (1 − p∗ )dc = 0.
From this expression, notice that if p < p∗ , then (recalling that dr < 0 and dc > 0) ∆(p) < 0
while if p > p∗ then ∆(p) > 0 by the same token.
With this in mind, shift things a little bit so that ∆(p∗ ) > 0 but ∆(p) < 0 if p ≤ p∗ − . Then
all types p below this threshold would still go to the single liability contract. Thus a tiny
change in the terms of the contract to a joint liability contract by one lender will completely
change the mix of the applaicants that he gets, generating positive profits for him. It is in
this way that an individual liability equilibrium is wiped out by the introduction of joint
liability contracts.
————————–
13.1.1 An Informal Extension to Entrepreneurship. To illustrate this last point, let us

redo and extend the two-skills example with private firms. What follows is just an informal
discussion with Suppose that production requires a setup cost S that must be borne up
front. Suppose that there are two wealth levels in the society, 0 and W (which exceed S).
Assume, moreoever, that there are just two productivity levels, λ = 0 and λ = 1, and that the
production function is given by λF(L), where L is the amount of labor hired. Finally, assume
that the high-productivity type can cover the setup cost of production.
In this economy, production will be carried out by those individuals for whom wealth equals
W and λ = 1. Let a be the fraction of such people. Note the following:
242 Credit Policy
(i) People with low wealth who are productive will not be able to bear the setup cost, and
this will be a loss to society. An increase in inequality that puts more people below the
threshold needed to finance setup costs will result in greater economic loss at the aggregate
level.
(ii) Assuming that the correlation between productivity and wealth is nonnegative, we see
that the loss is greatest when the correlation is zero (this is when a assumes its highest value).
The loss is minimal when productive types are only to be found among the wealthy (this
may be true if wealth is an indicator of past productivity, and if productivity exhibits high
serial correlation, both of which are strong assumptions).
(iii) The positive association between equality and greater efficiency depends critically on
the average wealth level in the society. If the average wealth level is high, then an increase
in equality puts more people above the threshold S needed to start up production. On
the other hand, if the average wealth level is low, then an increase in equality may throw
more individuals under the threshold, creating an even greater loss of efficiency. Thus poor
societies may be caught in a double-bind: where the (intrinsic) need for redistribution is
higher, the functional implications may be negative.
(iv) Denote the demand curve for labor by L(ν) (where the demand curve is obviously only
defined for the high types and where ν stands for the market wage). Then an increase in
inequality (assuming that we are not in the low average wealth trap) will lower a and move
the aggregate demand curve aL(ν) to the left. [So will a lowering of a caused by reduced
correlation between productivity and wealth.] If the wage rate is endogenous, it will go
down.
(v) This effect is heightened if those who cannot afford the setup costs to run a business
must join the labor force — the supply curve of labor shifts to the right at the same time
that the demand curve for labor shifts to the left. Effects (iv) and (v) are precisely the sort
of interactive implications highlighted by authors such as Banerjee and Newman [1993].
Inequality affects macroeconomic variables, such as the level of employment or the wage
rate.
(vi) Continuing this theme, we see that had we posited a slightly more complex model with a
continuum of productivity types, then productivity types which are supposed to be inactive
at the first best would enter production. This is because the equilibrium wage rate is low
relative to the first best, so that certain productivity types which would have not covered
the setup cost at the first best are now able to do so at the equilibrium wage rate. [These are
wealthy people as well.]
(vii) Adopting a slight political-economy perspective, we can see why coalitions of producers
might now want to oppose policies that provide better credit to the poor, even though these
policies may not be overtly redistributive.

10dnote Devraj PDF

Uploaded by

Copyright:

Available Formats

10dnote Devraj PDF

Uploaded by

Document Information

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

10dnote Devraj PDF

Uploaded by

Copyright:

Available Formats

Notes for a Course in Development Economics

Version 3.35, 2010.

Uneven Growth: A Research Agenda

The textbook paradigm of economy-wide development rests on the premise of “balanced

2.1 The Sources of Uneven Growth

2.1.1 The Composition of Demand. The textbook assumption of microeconomic theory

2.2 Reactions to Uneven Growth

2.2.1 Occupational Choice. Consider information-technology-based services in a devel-

2.2.2 Demand as a Channel for Spreading Income. A second avenue of percolation

2.3 Uneven Development: Final Remarks

The Calibration Game

3.1 Some Basic Facts

Low per capita incomes are an important feature of economic underdevelopment—perhaps

Figure 3.1. The Income Mobility of Countries, 1980–2000.

3.2 The Parente-Prescott and Lucas Calibrations

Normalize labor to 1, and consider the following aggregative production function:

In other words, the convergence predicted by technologically diminishing returns to inputs

Expectations and Multiple Equilibrium

4.2 Some Examples

4.2.4 Capital Deepening. Greater roundaboutness in production increases the produc-

4.2.8 Endogenous Growth. Economy-wide investment raises the return to individual

4.3 Complementarities and Development

4.4 Multiple Equilibrium

4.4.1 Complementarities through Demand. How do complementarities manifest them-

4.4.2 Pecuniary Externalities. Before I get into the Murphy-Shleifer-Vishny model it is

(4.6) ay(n) − F(n) = 0,

and rewriting this, we obtain the condition

Model 3. Separation of Costs and Benefits

4.4.5 Complementarity through the Production Process. “The division of labor is

(4.23) α = α(n1/(1−σ) ) ≡ A(n).

[3] A larger n affects final output (and therefore national income).

History Versus Expectations

In what follows, we look at some different approaches to the question of equilibrium-picking

5.1 State Variables and Common Knowledge

(5.2) h(x, ) + 2a(h(x, )) = x + 

Claim. If x∗ is some solution to (5.4), and x0 > x∗ , then

5.2 Applying Global Games to a Canonical Multiple Equilibrium Model

A canonical multiple equilibrium model can be described as follows. An economy consists

Per Capita Sectoral Return

Figure 5.1. A Depiction of Sectors A and B.

(5.8) r = f (K, z),

where f is strictly increasing in both K and z. As in the previous section z is a random

believe that the probability of having K people or more above me is precisely

(z +  − x̄) for z ∈ [x̄ − , x̄ + ]

5.3 Lagged Externalities

(5.12) ṙ(t) = g( f (K(t)) − r(t)).

This observation is independent of the magnitude of discounting, and of the degree of

5.4 State Variables and Time

Figure 5.2. Boundaries for Multiple Equilibria when z is Fixed.

k*(z) k(z,1) k(z,0)

Figure 5.3. Recursive Construction of the Locus k∗ (z).

k'(z,0) k'(z,1) k*(z)

Figure 5.4. Recursive Construction of the Locus k∗ (z).

6.2 Introductory Notes on Economic Inequality

6.3 A Special Note on Capital Stocks

Suppose that output is produced at each date t according to the equation

(U.1) The utility function u is increasing, continuous and strictly concave on R+ .

(6.5) F1 (k, k)G0 (0) ≤ β−1 − 1.

Then for any date T,

(5.2) h(x, ) + 2a(h(x, )) = x +

(z + − x̄) for z ∈ [x̄ − , x̄ + ]

||TM µ, TM µ0 || ≤ (1 − )||µ, µ0 ||,