On The Mechanics of Firm Growth: Erzo G. J. Luttmer

rdq028 2011/6/14 7:05 page 1042 #1
Review of Economic Studies (2011) 78, 10421068 doi: 10.1093/restud/rdq028

The Author 2011. Published by Oxford University Press on behalf of The Review of Economic Studies Limited.
Advance access publication 3 February 2011
On the Mechanics of Firm
Growth
ERZO G. J. LUTTMER
University of Minnesota and Federal Reserve Bank of
Minneapolis
First version received October 2008; nal version accepted June 2010 (Eds.)
The Pareto-like tail of the size distribution of rms can arise from random growth of productivity
or stochastic accumulation of capital. If the shocks that give rise to rm growth are perfectly correlated
within a rm, then the growth rates of small and large rms are equally volatile, contrary to what is found
in the data. If rm growth is the result of many independent shocks within a rm, it can take hundreds of
years for a few large rms to emerge. This paper describes an economy with both types of shocks that can
account for the thick-tailed rm size distribution, high entry and exit rates, and the relatively young age
of large rms. The economy is one in which aggregate growth is driven by the creation of new products
by both new and incumbent rms. Some new rms have better ideas than others and choose to implement
those ideas at a more rapid pace. Eventually, such rms slow down when the quality of their ideas reverts
to the mean. As in the data, average growth rates in a cross section of rms will appear to be independent
of rm size, for all but the smallest rms.
Key words: Firm size distribution, Gibrats law, Aggregate growth
JEL Codes: L1, O4
1. INTRODUCTION
Why does the employment size distribution of U.S. rms look like a Pareto distribution, with
the fraction of rms with more than n employees roughly equal to n
? Why is the tail in-

dex 1.05 barely high enough for the distribution to have a nite mean? More than half
of all rms with any employees have no more than four employees. But there are also almost
a 1000 rms with more than 10 000 employees each, and these rms employ as much as a
quarter of the U.S. labour force. What accounts for the large amount of heterogeneity in rm
size? How does this heterogeneity evolve over time? Some benchmark answers to these ques-
tions are needed for the systematic use of rm-level data in the study of aggregate growth and
uctuations.
In the presence of decreasing returns or downward sloping rm demand curves, it is possible
that the highly skewed size distribution entirely reects a highly skewed productivity distribu-
tion. Such a productivity distribution can arise if productivity growth is random and only suf-
ciently productive rms can survive. Given isoelastic cost functions or demand curves, random
productivity growth gives rise to Gibrats law, which holds that rm growth rates are indepen-
dent of size. A stationary size distribution results if employment at incumbent rms grows more
slowly on average than aggregate employment. This distribution has a tail index just above 1
1042

a
t

Q
u
e
e
n
'
s

U
n
i
v
e
r
s
i
t
y

o
n

J
u
l
y

8
,

2
0
1
4
h
t
t
p
:
/
/
r
e
s
t
u
d
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

rdq028 2011/6/14 7:05 page 1043 #2
LUTTMER ON THE MECHANICS OF FIRM GROWTH 1043
if cost parameters are such that there is only a small gap between entrant and incumbent mean
productivity growth rates (Luttmer, 2007).
1
This paper associates rm size not primarily with productivity differences, but with organi-
zation capital (Prescott and Visscher, 1980) that can be accumulated through investment over
time. In the model, a rm produces one or more differentiated commodities using labour and
commodity-specic blueprints. An entrepreneur can set up a new rm by producing a start-up
blueprint. After that, the rm can use labour and any of its blueprints to attempt to produce more
blueprints for new commodities. Individual blueprints can also become obsolete. The arrival
rates of these two types of events are independent and independent across blueprints. Absent
other sources of heterogeneity, this implies that the mean growth rate of a rm with more than a
single blueprint is independent of rm sizea weak version of Gibrats law. Averaging within
the rm implies that the variance of rm growth is inversely proportional to rm size, a violation
of the strong form of Gibrats law according to which the entire distribution of growth rates is
independent of rm size. The economy exhibits balanced growth, and increases in variety add to
the aggregate growth rate, as in Romer (1990) and Young (1998). As long as there is entry, the
size distribution will be stationary with a right tail that behaves like n
.
Independent within-rmreplication avoids a problemthat arises in economies with only rm-
wide productivity shocks. In Luttmer (2007), it takes a standard deviation of rm employment
growth of about 40% per annum to jointly account for the size distribution and the 11% rate of
rm entry observed in the data. This standard deviation is within the range reported by Davis
et al. (2007) for all rms, but implausibly high for large rms. Here, large rms are very stable
even when small-rm growth rates are sufciently volatile to be consistent with the observed
entry and exit rates. In the simplest version of the model, though, this is too much of a good
thing and leads to a rather dramatic counterfactual implication: the median age of rms with
more than 10 000 employees is implied to be about 750 years. Stationarity and the weak version
of Gibrats law force mean incumbent growth rates to be below the growth rate of the aggregate
labour force, only about 1% per annum, and averaging within the rm reduces variance by too
much for lucky rms to become large in a relatively short amount of time.
Newly collected data show that the median age of rms with more than 10 000 employees
in 2008 was only about 75 years. With a 40% standard deviation of employment growth, an
economy like Luttmer (2007) predicts about 100 years.
2
But to account for the relatively young
age of large rms observed in the data, without assuming there is a 30% chance that employment
at WalMart will grow or shrink by more than 40% over the next year, requires abandoning
Gibrats law.
Suppose therefore that some new rms enter with an initial blueprint of a higher quality than
other blueprints in the economy. The resulting higher prots per blueprint create an incentive
to copy these blueprints at a higher rate if quality is inherited. If copies stay within the rm,
then these new rms will grow fast. If a rms quality advantage is transitory, this rapid growth
will come to an end eventually. A stationary distribution with a tail index above 1 results if
there is positive entry along the balanced growth path. A simple formula shows that this tail
index will be close to 1 if rms with high-quality blueprints grow at an equilibrium rate that is
1. The =1 asymptote is known as Zipfs law. See Axtell (2001) for recent evidence on the rm size distribution
showing that slightly above 1 ts the data well. Well-known empirical studies on Gibrats law for rms, based on
growth rate regressions that correct for selection, are Evans (1987) and Hall (1987). Sutton (1997) surveys the literature.
Gabaix (1999) uses Gibrats law to interpret the city size distribution and contains many useful references on the history
of the subject. Rossi-Hansberg and Wright (2007) develop a model of the rm size distribution in which there are many
industries and the rm size in any given industry follows a stationary process instead of the non-stationary process
implied by Gibrat.
2. A new calibration is available at www.luttmer.org.

a
t

Q
u
e
e
n
'
s

U
n
i
v
e
r
s
i
t
y

o
n

J
u
l
y

8
,

2
0
1
4
h
t
t
p
:
/
/
r
e
s
t
u
d
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

rdq028 2011/6/14 7:05 page 1044 #3
1044 REVIEW OF ECONOMIC STUDIES
FIGURE 1
Selected growth histories
slightly below the sum of the growth rate of the aggregate labour force and the hazard rate with
which high-quality rms lose their edge. Thus, high-quality rms can grow fast if the period of
rapid growth is not expected to last too long. But there will be variation in how long rms are
in this rapid growth phase, and this variation allows for the appearance of young large rms.
This version of the organization capital interpretation of rm growth can match the overall size
distribution, the amount of entry and exit, as well as the relatively young age of large rms.
Furthermore, although Gibrats law does not hold, the mean growth rates of surviving rms
behave like they do in the data: roughly independent of size for most rms and signicantly
higher for the smallest rms (Dunne, Roberts and Samuelson, 1989).
Figure 1 presents some corroborating evidence for the type of histories of rm growth pre-
dicted by the model. It shows the employment histories of 25 of the nearly 1000 large rms that
had more than 10 000 employees in 2008 (the data are described in Appendix A). The average
employment growth rate across all rms reported in Figure 1 is almost 18% per annum, and there
is considerable variation. In particular, rm growth rates seem to be much above average when
rms are relatively small and decline signicantly when rms become large. The data shown
in Figure 1 represent only a small sample from a population of slightly under a thousand large
rms. In turn, this population of large rms was selected over many years from the population
of all rms that were ever set up. In U.S. data, the number of rms grows at an annual rate about
equal to the 1% growth rate of aggregate employment. Combined with an entry rate of 11%, a
steady-state calculation implies that the number of rms that was ever set up is roughly 11 times
the current population of around 6 million rms.
3
The thousand or so rms with ten thousand or
more employees are thus a highly selected sample from a universe of about 66 million rms. In
such a selected sample, one might conjecture, it is not surprising to see that large rms tend to
3. The growth rate of the collection of all historical rms equals the entry rate times the fraction of all historical
rms that are currently active. In a steady state, it also equals the growth rate of the active number of rms, which equals
the growth rate of aggregate employment.

a
t

Q
u
e
e
n
'
s

U
n
i
v
e
r
s
i
t
y

o
n

J
u
l
y

8
,

2
0
1
4
h
t
t
p
:
/
/
r
e
s
t
u
d
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

rdq028 2011/6/14 7:05 page 1045 #4
have a history of rapid enough growth to match the age distribution, even though Gibrats law
holds. The results presented in this paper show that this conjecture is wrong when shocks tend to
average out within a rm. The strings of positive growth needed are too unlikely, and currently
active large rms should be about 750 years old if Gibrats law holds.
1.1. Related literature
This paper goes back to, interprets, and builds on the type of growth process initially proposed
by Yule (1925) and Simon (1955). Yule (1925) was concerned with the number of species in
biological genera, and Simon (1955) with word frequencies, city sizes and income distributions.
In the context of cities, (Krugman, 1996, p. 96) described the time it takes for cities to grow large
in Simons model as an unresolved problem. Simon and Bonini (1958), Ijiri and Simon (1964)
and many others since studied rm growth. Klette and Kortum (2004) describe an economy
based on the quality-ladder model of Grossman and Helpman (1991) in which rm size follows
a birthdeath process, as in this paper. In their economy, incumbent rms cannot growon average
because there is a xed set of commodities and new entrants continuously capture the markets
for some of those commodities. This makes it impossible for large rms to arise. This difculty
is resolved here by considering an economy in which the number of commodities can grow over
time, as in Romer (1990) and Young (1998). Even without growth in the number of markets, a
thick-tailed size distribution can arise in the Klette and Kortum (2004) economy if Gibrats law
is relaxed along the lines described in this paper.
The models in this paper are highly tractable analytically and inevitably stylized. Lentz and
Mortensen (2006) use a version of the Klette and Kortum (2004) economy with additional and
more exible sources of heterogeneity. They do not address the thin-right-tail problem but esti-
mate their model using panel data on Danish rms.
4
The Danish rm size data do not appear to
exhibit the striking Pareto shape that is found reliably in U.S. data. The small size of the Danish
economy may well account for thisthere are as many rms in the U.S. as there are people in
Denmark. When it comes to examining the right tail of the size distribution, a model economy
with a continuum of rms could simply be a better abstraction for the U.S. than for a small coun-
try like Denmark. In addition, small countries will have fewer very large rms if the replication
of blueprints across national boundaries or outside language areas comes at additional costs.
Firms in this paper are organizations that operate in (monopolistically) competitive markets
and grow through continuous investment in new blueprints, at a level that is proportional to
the size of the rm. One can alternatively view a rm as a trading post or network in which
agents trade repeatedly. Gibrats lawand the observed size distribution arise if there is population
growth and agents search for rms by randomly sampling other agents and matching with the
rm with which the agent sampled is already matched. A simple version of such a model is
described in Luttmer (2006). Related models of network formation are presented in Jackson
(2006) and Jackson and Rogers (2007), and the extensive literature cited therein. Deciding on
the relative importance of these alternative interpretations poses difcult identication problems.
1.2. Outline
The economy and its balanced growth path are described in Section 2, together with two alterna-
tive formulations of the role of blueprints in production. The stationary size and age distributions
are derived in Section 3 and formulas are given for the tail index in the Gibrat and non-Gibrat
cases (Propositions 3 and 4), and for the mode of the age distribution of large rms when both
4. See also Seker (2007) for related work on Chilean establishments.

a
t

Q
u
e
e
n
'
s

U
n
i
v
e
r
s
i
t
y

o
n

J
u
l
y

8
,

2
0
1
4
h
t
t
p
:
/
/
r
e
s
t
u
d
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

rdq028 2011/6/14 7:05 page 1046 #5
Gibrats and Zipfs law hold (Section 3.5). Calibrations are in Section 4. All proofs and a de-
scription of the data are in the appendix.
2. THE ECONOMY
Blueprints are costly to replicate or produce fromscratch. In the baseline version of the economy,
a blueprint describes the idea for a particular nal good. No two blueprints are the same, whether
produced by replication or from scratch. Final goods producers are monopolistic competitors.
Time is continuous and indexed by t [0, ).
2.1. Consumers
There is a growing population of consumers measured by H
t
= He
t
at time t . The dynastic pref-
erences of the representative consumer over aggregate consumption ows C
t
are determined by
E
0
__

0
e
t
H
t
(C
t
/H
t
)
1
1
dt
_
.
The parameters , and are positive and = 1 is interpreted as logarithmic utility. Mar-
kets are complete and consumers face standard budget constraints. The resulting interest rate in
consumption numeraire is related to the consumption growth rate via
r
t
= +
_
DC
t
C
t
_
. (1)
Aggregate consumption is a CES composite of differentiated commodities, as in Dixit and
Stiglitz (1977),
C
t
=
__
C
11/
,t
dN
t
()
_
1/(11/)
,
where N
t
() is the measure of type- commodities and > 1 is the elasticity of substitution. In
the baseline specication, all commodity types are the same and all producers choose to charge
the same price p
t
. Consumers therefore set C
,t
=C
t
( p
t
), which implies that
C
t
=C
t
( p
t
)N
1/(11/)
t
, (2)
where N
t
=
_
dN
t
(). Cost minimization implies that commodity demands are
C
t
( p) =
_
p
P
t
_
C
t
, (3)
where P
t
is the price index P
t
= p
t
N
1/(1)
t
. Note that the prices of differentiated commodi-
ties and the composite good are quoted in some arbitrary numeraire. All other prices will be
expressed in units of the composite commodity.
2.2. Producers
Given a blueprint for a particular differentiated commodity, a producer can use l units of labour
to produce Z
t
l units of the commodity, where Z
t
= Ze
t
evolves exogenously. Given a real
wage w
t
, the marginal cost of one unit of a commodity is thus w
t
/Z
t
in units of composite
good, and the constant elasticity demand curves (3) imply that producers set prices at a constant

a
t

Q
u
e
e
n
'
s

U
n
i
v
e
r
s
i
t
y

o
n

J
u
l
y

8
,

2
0
1
4
h
t
t
p
:
/
/
r
e
s
t
u
d
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

rdq028 2011/6/14 7:05 page 1047 #6
markup over marginal cost, p
t
/P
t
=(w
t
/Z
t
)/(11/). Combining this with equations (2) and
(3) determines the equilibrium real wage as a function of the state (Z
t
, N
t
),
w
t
=(11/)Z
t
N
1/(1)
t
. (4)
The amount of labour needed to satisfy the resulting demand for a typical commodity is l
t
=
C
t
( p
t
)/Z
t
. Using equation (3) and the prices p
t
/P
t
set by producers, this gives
l
t
=
_
(11/)Z
t
w
t
_
1
(11/)C
t
w
t
. (5)
The markup 1/(11/) of price over marginal cost implies that prots measured in units of the
composite good will be w
t
l
t
/( 1).
2.3. New blueprints
The producer of a differentiated commodity needs a blueprint to produce. Blueprints depreciate
in a one-hoss-shay fashion at an average rate
t
. New blueprints for distinct differentiated com-
modities can be produced by using labour to replicate existing blueprints, or from scratch by
entrepreneurs. The respective rates at which this occurs in equilibrium are denoted by
t
and
t
(a mnemonic for
t
,
t
and
t
is less, more and new.) The number of blueprints therefore
evolves according to
DN
t
=(
t
+
t
t
)N
t
. (6)
An initial condition determines N
0
.
2.3.1. Replication of existing blueprints. A new blueprint produced from an existing
blueprint arrives following an exponentially distributed waiting time with mean
t
= f (i
t
),
where i
t
is labour employed in the replication process. An existing blueprint is lost follow-
ing an exponentially distributed waiting time with mean
t
= g( j
t
), where j
t
is labour used to
maintain the blueprint. Note that an existing blueprint generates revenues from its use in the
production of a commodity and as an input in the production of new blueprints.
5
The value q
t
of a blueprint must satisfy the Bellman equation
r
t
q
t
= max
f (i )
g( j )
_
w
t
_
l
t
1
[i + j ]
_
+()q
t
+Dq
t
_
, (7)
together with a transversality condition. The blueprint production function f is increasing and
exhibits strictly decreasing returns to scale. The blueprint depreciation function g is assumed to
be strictly decreasing and convex. For simplicity, both f and g are assumed to be sufciently
smooth, with slopes that are unbounded near zero and converge to zero for large i and j . The
optimal levels of investment in new blueprints and maintenance of existing blueprints are deter-
mined by
t
= f (i
t
),
t
= g( j
t
), q
t
D f (i
t
) =q
t
Dg( j
t
) =w
t
. (8)
The technology assumptions ensure that
t
and
t
are increasing in q
t
. Blueprints are repli-
cated more quickly and maintained better when their value is high.
5. The model of how Wal-Mart has expanded since 1962 described in Holmes (2006) has this feature. The key
assumption here is that K-Mart cannot simultaneously look at a Wal-Mart blueprint to produce a new blueprint of its
own. As in Boldrin and Levine (1999, 2006), and unlike Luttmer (2007), spillovers are assumed to be of secondary
importance in this economy.

a
t

Q
u
e
e
n
'
s

U
n
i
v
e
r
s
i
t
y

o
n

J
u
l
y

8
,

2
0
1
4
h
t
t
p
:
/
/
r
e
s
t
u
d
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

rdq028 2011/6/14 7:05 page 1048 #7
2.3.2. Newdesigns by entrepreneurs. Newblueprints can also be designed fromscratch
by agents acting as entrepreneurs, without the input of an existing blueprint. At any point in time,
every agent in the economy is endowed with one unit of effort that can be allocated to two tasks:
supplying labour or attempting to produce a blueprint. Every agent has a skill vector (x, y),
where x is the rate at which the agent can develop new blueprints and y is the amount of labour
the agent can supply per unit of time. Comparative advantage determines occupational choice.
Ignoring ties, agents with skill vectors that satisfy q
t
x > w
t
y will choose to be entrepreneurs
who design blueprints and agents with skill vectors that satisfy q
t
x < w
t
y will choose to be
employees.
There is a time-invariant talent distribution T dened over the set of all possible skill vectors,
as in the Roy model of Rosen (1978). This talent distribution has a nite mean. For simplicity,
it is assumed to have a density so that ties play no role. The resulting per capita supplies of
entrepreneurial effort and labour are
E(q
t
/w
t
) =
_
q
t
xw
t
y
xdT(x, y), (9)
L(q
t
/w
t
) =
_
q
t
x<w
t
y
ydT(x, y), (10)
respectively. Clearly, the supply of entrepreneurial effort is increasing in q
t
/w
t
, and the supply
of labour is decreasing, both ranging between 0 and the mean skill in the population. If the
talent distribution is Frchet, then the elasticity of E(q
t
/w
t
)/L(q
t
/w
t
) with respect to q
t
/w
t
is
constant (Luttmer, 2008).
2.4. Equilibrium
Given a per capita supply of entrepreneurial effort E(q
t
/w
t
) and a stock of blueprints N
t
, the
rate
t
at which new blueprints are added by entrepreneurs is determined by
t
N
t
= H
t
E(q
t
/w
t
). (11)
Labour market clearing requires that
(l
t
+i
t
+ j
t
)N
t
= H
t
L(q
t
/w
t
). (12)
The equilibrium is determined by equations (1)(12), an initial condition for N
0
and a transver-
sality condition for q
t
N
t
.
Because the product market distortion arising from monopolistic competition is the same in
all markets and at all times, and because agents supply their time inelastically, it turns out that the
equilibrium allocation is Pareto efcient. It is possible to characterize the equilibrium dynamics
in terms of only one state and one costate variable and use a phase diagram to construct an
equilibrium that converges over time to a balanced growth path.
6
2.5. Balanced growth
A key feature of the balanced growth path is that the allocation of labour per blueprint is con-
stant at some (i, j,l). Population growth then implies that the measure of blueprints is given by
6. The rate at which blueprint capital is accumulated in this economy depends intricately on the shape of the
production and depreciation functions f and g and the shape of the talent distribution. Adjustment to the balanced
growth path may be slow and asymmetric. A detailed analysis is beyond the scope of this paper.

a
t

Q
u
e
e
n
'
s

U
n
i
v
e
r
s
i
t
y

o
n

J
u
l
y

8
,

2
0
1
4
h
t
t
p
:
/
/
r
e
s
t
u
d
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

rdq028 2011/6/14 7:05 page 1049 #8
N
t
= Ne
t
for some N. Because of equations (4) and (5), wages and per capita consumption
grow at the rate = +/( 1). The term /( 1) measures gains from variety, as in Young
(1998). By equation (1), the implied interest rate is r = + . The ow prot from producing
a commodity is w
t
l/( 1). It follows that [q
t
, w
t
] =[q, w]e
t
. The Bellman equation (7) then
implies that wages and blueprint prices must satisfy the present-value condition
q
w
=
l
1
(i + j )
r ()
, (13)
where (i, j ) and (, ) satisfy
= f (i ), = g( j ), (q/w)D f (i ) =(q/w)Dg( j ) =1. (14)
Holding xed l, these conditions imply that q/w is equal to the maximum subject to [, ] =
[ f (i ), g( j )] of the right-hand side of equation (13), as long as this is nite.
7
The fact that the
aggregate number of blueprints grows at the rate implies that new blueprints must be added by
entrepreneurs at the rate = (). If E(q/w) is positive, then the entrepreneurial supply
of blueprints (11) determines the steady-state supply of blueprints via
N
H
=
E(q/w)
()
. (15)
Alternatively, E(q/w) = 0 and = . Along a balanced growth path, the labor-market
clearing condition (12) implies a derived demand for blueprints equal to
N
H
=
L(q/w)
i + j +l
. (16)
The balanced growth conditions (13)(16) determine (i, j,l), (, ), q/w and N/H. The level
of wages follows from (4) and aggregate consumption can be obtained from equation (5), wages,
and l.
Given a positive q/w that is not too large, the conditions (13)(14) can be solved for the
labour allocation (i, j,l) and the resulting blueprint creation and destruction rates and . It is
not difcult to verify that (i, j,l) and are increasing in q/w. Since E(q/w) is increasing
in q/w, this implies a steady-state supply of blueprints (15) that is increasing in q/w. Since
L(q/w) is decreasing in q/w, the derived demand for blueprints (16) is decreasing in q/w.
There can therefore be at most one price q/w that clears the market for blueprints in steady
state.
The replication technology must be assumed to satisfy f (0) g(0) < or else
cannot hold. The assumption r > ensures that implies r > . The
fact that E(q/w) and L(q/w) go to zero as q/w goes to, respectively, zero and innity, can
now be used to argue that equations (13)(16) does in fact have a solution. It remains to show
that the decision problem of blueprint owners is well dened. This follows because blueprint
owners cannot obtain unbounded prots by replicating more quickly than r . The fact that
f (i ) g( j ) is increasing and concave implies that, at the proposed equilibrium, the ow cost
of doing so would exceed the ow revenues wl/( 1) per blueprint. Together, these results
establish the following proposition.
7. That is, if and only if l is low enough to ensure that max
i, j
{f (i ) g( j ) : i + j l/( 1)} does not exceed
r . The value of a blueprint is innite for l outside this range.

a
t

Q
u
e
e
n
'
s

U
n
i
v
e
r
s
i
t
y

o
n

J
u
l
y

8
,

2
0
1
4
h
t
t
p
:
/
/
r
e
s
t
u
d
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

rdq028 2011/6/14 7:05 page 1050 #9
Proposition 1. Suppose that + > + and > f (0) g(0). Suppose that the talent
distribution is such that E(q/w) > 0 for all strictly positive q/w. Then, equations (13)(16)
dene the unique balanced growth path and > .
Abalanced growth path with E(q/w) =0 can arise if the talent distribution has bounded support.
In such an equilibrium, new blueprints are only produced using replication from an initial stock
of blueprints.
2.6. Alternative blueprint interpretations
In the setup considered so far, different blueprints specify distinct differentiated commodities
that are produced subject to constant returns and are sold to all consumers. The equilibrium con-
ditions for this economy also apply to an economy in which consumers live in many different
locations and blueprints are location specic. With minor modications, the same framework can
be used as well to consider competitive nal goods markets and blueprints containing the spec-
ications for production facilities or plants that are subject to decreasing returns. The following
discussion elaborates on these two interpretations. They are benchmarks. Hybrid formulations
are more plausible, but also less analytically tractable.
2.6.1. Sales ofces or stores. Suppose that at any point in time, consumers are evenly
distributed across many locations. In each location, there are many consumers who can only buy
from local stores. Preferences are as in equation (2), with N
t
now denoting the measure of stores
in a particular location. An entrepreneur can create a blueprint for a store in a randomly selected
location. The store sells a new differentiated product. The blueprint can then be copied to operate
stores selling the same differentiated product in randomly selected new locations. There is an
economy-wide market for labour services, or, equivalently, output is produced where workers
live and can be shipped to stores at no cost.
Because there are many locations, replicated blueprints are always assigned to new locations,
and every new store sells a commodity that is new to the market in which it is introduced.
Assuming that there is a very large number of blueprints that can be copied, every location
receives a constant ow of new stores, and stores are uniformly distributed across locations.
As a result, new stores face the same market conditions everywhere.
8
With this, the analysis
proceeds as before.
2.6.2. Jobs, production lines, plants. Instead of assuming that the output of every pro-
ducer is unique, suppose there is one competitive market for nal goods. A blueprint denes
a particular job, production line or plant that is subject to decreasing returns to scale. Each
plant can use l
t
units of labour to produce output Z
t
F(1,l
t
) for some constant returns to scale
production function F. Growth in variety is no longer a source of consumption growth. Along a
balanced growth path, wages growat the same rate as Z
t
, the amount of labour used per blueprint
is constant, and the number of blueprints grows at the population growth rate . In contrast to the
DixitStiglitz formulation used elsewhere in this paper, no constant elasticity assumptions are
8. There must be many more stores than locations. Imagine markets are non-overlapping intervals of length 1/A
in [0, 1], where A N. Each one of the A markets has A consumers and there are A
2
stores that are randomly assigned
to points in [0, 1]. The ratio of stores to consumers is /. As A becomes large, the proportion of all stores assigned
to the region [0, x] converges to x. If the number of stores were A instead, then the number of stores in different
markets would remain random and converge to a Poisson distribution. Market conditions would vary across locations,
and strategic considerations would come into play in each market.

a
t

Q
u
e
e
n
'
s

U
n
i
v
e
r
s
i
t
y

o
n

J
u
l
y

8
,

2
0
1
4
h
t
t
p
:
/
/
r
e
s
t
u
d
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

rdq028 2011/6/14 7:05 page 1051 #10
needed. The production function F is general, even though Z
t
is not labour augmenting in the
usual sense. In units of blueprints, the per capita capital stock is constant. But the market value
of the capital stock and the cost of producing new capital grows at the same rate as Z
t
, wages,
and output per capita.
3. THE DISTRIBUTION OF FIRM SIZE AND AGE
The economy described up to now has agents who consume, supply labour, and act as en-
trepreneurs. Everyone can own blueprints and there are no rms. A transaction cost argument
can be used to motivate a denition of what rms are in this economy.
Consider an entrepreneur who has just developed a new blueprint. To hire labour to produce
the associated commodity and develop further copies of the same blueprint, the entrepreneur can
set up a rm at no cost. This denes a rm entry. Claims to rms can be traded freely. But there
is a cost, potentially very small, involved in rms hiring entrepreneurs to develop new blueprints
from scratch, in selling blueprints to rms, and in merging rms. There are no cost advantages
to any of these transactions, and so they will not occur in equilibrium.
9
A rm will therefore
only gain new blueprints through organic growth by replicating its existing blueprints. A rm
only loses blueprints as they depreciate at the rate .
10
Exit occurs when a rm has lost all its
blueprints.
The measure of rms with n blueprints at time t is denoted by M
n,t
. The aggregate measure
of blueprints is therefore
N
t
=
n=1
nM
n,t
. (17)
Over time, the change in the number of rms with one blueprint is
DM
1,t
=2M
2,t
+N
t
(+)M
1,t
, (18)
where , and =() are equilibrium rates that are constant along the balanced growth
path. The number of rms with one blueprint increases because rms with two blueprints lose
one or because of entry. The number declines because rms with one blueprint gain or lose a
blueprint. Similarly, the numbers of rms with more than one blueprint evolve according to
DM
n,t
=(n 1)M
n1,t
+(n +1)M
n+1,t
(+)nM
n,t
, (19)
for all n 1 N. The joint dynamics of N
t
and {M
n,t
}
n=1
is fully described by equations (17)
(19).
3.1. The stationary size distribution
Along the balanced growth path, N
t
grows at the rate and a stationary rm size distribution
exists if equations (17)(19) have a solution that satises DM
n,t
= M
n,t
for all n N. Given
that N
t
and M
n,t
grow at the common rate , one can then dene
P
n
=
M
n,t
k=1
M
k,t
9. Of course, these transactions do occur in the data. This is a familiar and important failure of the type of model
described in this paper. Chatterjee and Rossi-Hansberg (2006) provide an interesting model of rm size in which adverse
selection makes it difcult for rms to hire entrepreneurs.
10. Bernard, Redding and Schott (2006) document the importance of turnover in the mix of products sold by U.S.
manufacturing rms. They report that less than 1%of product adds and drops are associated with mergers or acquisitions.

a
t

Q
u
e
e
n
'
s

U
n
i
v
e
r
s
i
t
y

o
n

J
u
l
y

8
,

2
0
1
4
h
t
t
p
:
/
/
r
e
s
t
u
d
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

rdq028 2011/6/14 7:05 page 1052 #11
for all n N. This is the fraction of rms with n blueprints. It is analytically more convenient to
use the fraction of all blueprints held by rms of size n. This is given by
Q
n
=
nM
n,t
k=1
kM
k,t
for all n N. With these denitions, equation (18) becomes
Q
1
=Q
2
+ (+)Q
1
, (20)
and equation (19) implies
1
n
Q
n
=Q
n1
+Q
n+1
(+)Q
n
, (21)
for n 1 N. Note that these equations only depend on the parameters / and /. The
stationary distribution cannot depend on the units in which time is measured.
Proposition 2. Suppose that , , and = () are positive. Dene the se-
quence {
n
}
n=0
by the recursion
n
= 1/(1 (
n1
/) +( +n)/n) and the initial con-
dition
0
=0. This sequence is monotone and converges to min{1, /}. The only non-negative
and summable solution to equations (20)(21) is given by
Q
n
=

k=0
1
n+k
_
n+k
m=n
m
_
n+k
m=1
. (22)
If =, then
Q
n

||
n1
m=1
. (23)
Here, equation (23) means that the ratio of the left- and right-hand sides converges to 1 as n
becomes large. If = 0, then the only non-negative and summable solution to equations (20)
(21) is identically zero, implying that there does not exist a stationary in this case. If > 0, then
equation (22) adds up to 1 by construction and denes a stationary size distribution {P
n
}
n=1
via
P
n
Q
n
/n. The mean rm size can be written as 1/(
n=1
Q
n
/n), and this is also nite by
construction. Appendix B proves these results and gives an explicit solution for Q
n
in the more
general case that arises when the size distribution of entrants is non-degenerate.
When > , the properties of the right-hand side of equation (23) are very different from
what they are when >. If >, then Q
n
is bounded above by a multiple of the geometrically
declining sequence (/)
n
. On the other hand, if > , then
n
/ 1, and hence the right-
hand side of equation (23) declines at a rate that is slower than any given geometric rate. The
following proposition gives a further characterization of the right tail of the distribution.
Proposition 3. Suppose that > >0. Then, the right tail probabilities R
n
=
k=n
P
k
of the stationary rm size distribution satisfy
lim
n
n
_
1
R
n+1
R
n
_
=,
where =/(). That is, R
n
is a regularly varying sequence with index .

a
t

Q
u
e
e
n
'
s

U
n
i
v
e
r
s
i
t
y

o
n

J
u
l
y

8
,

2
0
1
4
h
t
t
p
:
/
/
r
e
s
t
u
d
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

rdq028 2011/6/14 7:05 page 1053 #12
This means that lim
n
R
[xn]
/R
n
= x
for any x > 0, where [xn] is the integer part of xn.

An implication is that n
z
R
n
0 for all z < and n
z
R
n
for all z > . Even though this
does not describe precisely what happens to n
R
n
for large n, the parameter will continue to
be referred to as the tail index of the size distribution.
11
The limiting tail index =1 associated
with Zipfs law arises when the rate = () at which blueprints are introduced by
entrepreneurs converges to zero. Appendix B shows that Proposition 3 holds more generally if
the size distribution of entrants has a right tail that is regularly varying with an index smaller
than .
For comparison, consider the economy of Klette and Kortum (2004). There, =0 and <.
This turns equations (20)(21) into a linear difference equation with constant coefcients that is
easy to solve. The resulting rm size distribution is R.A. Fishers logarithmic series distribution,
which has P
n
(/)
n
/n. As a result, right tail probabilities converge to zero even more quickly
than a geometric sequence. To generate a thick right tail, rms must grow on average, and in the
economy described here this requires population growth. A tail index close to 1 can only arise
if growth in the number of blueprints is mostly due to incumbents rather than new entrants. It is
critical that rms grow exponentially. If rms accumulate new blueprints at some constant rate
, instead of n, then the size distribution would be Poisson like, with a geometrically bounded
right tail.
3.2. Firm entry and exit rates
The ow of blueprints introduced by new rms is N
t
. Each new rm starts with one blueprint,
and so N
t
is also the ow of new rms that enter per unit of time. The rm entry rate is therefore
= N
t
/M
t
, where M
t
=
n=1
M
n,t
is the number of rms in the economy. It follows that
/ = N
t
/M
t
=
n=1
nP
n
is the average rm size. An alternative way to calculate the rm
entry rate is to note that the only rms that exit in this economy are rms with one remaining
blueprint. The proportion of such rms is P
1
, and they exit at a rate . The resulting balance
P
1
of rms entering and exiting per unit of time must equal the rate at which the number
of rms grows over time. These two calculations can be summarized as
=
n=1
nP
n
= +P
1
. (24)
Just like the blueprint entry rate + = + , the rm entry rate can be no less than the
population growth rate, and this lower bound is attained only when rms never lose blueprints
and therefore never exit. The two equations given in equation (24) and Q
1
= P
1
/
n=1
nP
n
imply / = /( Q
1
). Together with equation (22), this yields an explicit formula for the
rm entry rate relative to the population growth rate. In turn, this implies an explicit formula for
the average rm size /.
3.3. Firm type transitions
The data shown in Figure 1 suggest that some rms initially grow at rates that far exceed the
bound < implied by Proposition 1 and that these growth rates decline with rm size and
age. A simple way to account for this slowdown and examine its implications for the stationary
size distribution is as follows. Suppose there are high- and low-quality blueprints. High-quality
blueprints imply a productivity Z
H
e
t
and low-quality blueprints imply a productivity Z
L
e
t
,
11. See Bojanic and Seneta (1973) for the denition of regularly varying sequences and some of its implications.
Bingham, Goldie and Teugels (1987) is a useful source on the general topic of regular variation.

a
t

Q
u
e
e
n
'
s

U
n
i
v
e
r
s
i
t
y

o
n

J
u
l
y

8
,

2
0
1
4
h
t
t
p
:
/
/
r
e
s
t
u
d
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

rdq028 2011/6/14 7:05 page 1054 #13
where Z
H
> Z
L
. Entrepreneurs produce high-quality blueprints with probability (0, 1] and
low-quality blueprints with probability 1 .
12
Incumbent rms replicate blueprints as before,
preserving their quality. But high-quality rms transition to become low-quality rms following
independent and exponentially distributed waiting times with a mean 1/
H
. When such a transi-
tion happens, all blueprints of the rm turn into low-quality blueprints, permanently. Any new
blueprints created by the rm thereafter will be of low quality. Low-quality rms can also exit
randomly at a rate
L
, irrespective of the number of blueprints that make up the rm.
One possible interpretation for these kinds of rm type transitions is that some aspect of the
environment for which the initial blueprint of a rm was created changes permanently. Outside
the formal model described here, a relative decline in the quality of a rms blueprints could arise
from competing rms catching up. An alternative interpretation for the decline in rm growth
rates is that blueprints are location specic and that rms initially implement blueprints in the
most protable locations.
Along a balanced growth path, the present-value condition (13) must be modied to account
for the loss in value that occurs when a blueprint transitions from one type to another. The
incentives to replicate and maintain continue to be determined by equation (14). Let q =q
H
+
(1 )q
L
, where q
H
and q
L
are the respective prices of high- and low-quality blueprints. The
steady-state number of high-quality blueprints is N
H
/H = E(q/w)/( +
H
[
H
H
]).
Low-quality blueprints are produced by entrepreneurs, by incumbent replication, and because a
ow of high-quality blueprints depreciate in quality. This implies N
L
/H = [(1 )E(q/w) +
H
N
H
/H]/(+
L
[
L
L
]). Write (i
I
, j
I
,l
I
) for the allocation of labour to a type-I blueprint.
As long as the talent distribution for entrepreneurs is unbounded, E(q/w) is positive and the
labour market clears if
L(q/w)
E(q/w)
=
(1)(i
L
+ j
L
+l
L
)
+
L
(
L
L
)
+
+
H
(
H
H
)
_
i
H
+ j
H
+l
H
+

H
(i
L
+ j
L
+l
L
)
+
L
(
L
L
)
_
.
As in the case of equations (15)(16), labour market clearing forces +
H
>
H
H
and
+
L
>
L
L
in any equilibrium in which entrepreneurs contribute to the supply of new
blueprints.
One can verify that Z
H
> Z
L
implies i
H
> i
L
, j
H
> j
L
, l
H
> l
L
and q
H
> q
L
. While their
quality advantage lasts, high-quality rms have stronger incentives to replicate and maintain
blueprints than low-quality rms. High-quality rms choose to grow faster than low-quality
rms. This can account for the thick tail of the size distribution.
Proposition 4. Suppose some rms enter as high-quality rms and transition to low-quality
rms at a positive rate
H
. Low-quality rms also exit randomly at a rate
L
. Then, along the
balanced growth path,
H
H
>
L
L
, +
H
>
H
H
and +
L
>
L
L
. The
stationary size distribution has a tail index given by
=min
_
+
H
[
H
H
]
+
,
+
L
[
L
L
]
+
_
.
The right tail of the size distribution declines geometrically if this is innite.
12. Alternatively, one can assume that entrepreneurs have potentially different skills for producing high- and low-
quality blueprints. Relative prices and comparative advantage then determine the quality mix of start-up blueprints.

a
t

Q
u
e
e
n
'
s

U
n
i
v
e
r
s
i
t
y

o
n

J
u
l
y

8
,

2
0
1
4
h
t
t
p
:
/
/
r
e
s
t
u
d
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

rdq028 2011/6/14 7:05 page 1055 #14
The actual size distribution and a proof of Proposition 4 are implied by the results in Appendix
B. If = ( +
H
)/(
H
H
), then large rms arise because of the rapid growth of new rms.
This can generate a thick tail even if there is no population growth.
3.4. Firm age and size
The age distribution among large rms is a useful tool for assessing alternative interpretations
of the rm size distribution. This age distribution can be constructed from the size distributions
of cohorts of rms that are the same age.
3.4.1. The size distribution of a cohort. Consider a cohort of rms initially with k N
blueprints and in a common growth phase. While in this initial growth phase, rms of size n N
gain blueprints at the rate n and lose blueprints at the rate n. Firms in this cohort also transition
randomly into a new growth phase at a rate . Let T
1,k
(a) denote the fraction of rms in the
cohort that have made this transition by age a. Dene T
n,k
(a) to be the fraction of rms that
have n N blueprints at age a and have not made the transition. Let T
0,k
(a) represent the rms
in the cohort that have exited as a result of losing their last blueprint. Only rms that have not
yet exited can transition into a new growth phase,
DT
1,k
(a) =[1T
1,k
(a) T
0,k
(a)]. (25)
Exit occurs when a rm loses its last blueprint, and hence
DT
0,k
(a) =T
1,k
(a). (26)
The number of rms with n blueprints and still in the original growth phase by age a must satisfy
DT
n,k
(a) =(n 1)T
n1,k
(a) +(n +1)T
n+1,k
(a) [ +(+)n]T
n,k
(a) (27)
for all n N. Note that the T
n,k
(a) term is not scaled by n, reecting the assumption that the
transition to a new growth phase is independent of size.
Proposition 5. For any > 0 and 0 dene (a) =(e
()a
1)/(e
()a
/). Fix
some k N. Then, equations (25)(27) together with the initial condition T
k,k
(0) =1 give
T
1,k
(a) =
_
a
0
e
b
_
1
_
(b)
_
k
_
db
and
T
0,k
(a) =k
_
a
0
e
b
_
1

(b)
__
(b)
_
k1
[1 (b)]db,
as well as
T
n,k
(a) =e
a
min{n,k}
m=1
_
k
m
__
n 1
m1
__
1

(a)
_
m
_
(a)
_
km
[1 (a)]
m
nm
(a),
for all n N.
For = 0 and k = 1, this solution can be found in Klette and Kortum (2004). The probability
generating function for =0 and k Nis in Kendall (1948). Using the fact that (a) goes to zero

a
t

Q
u
e
e
n
'
s

U
n
i
v
e
r
s
i
t
y

o
n

J
u
l
y

8
,

2
0
1
4
h
t
t
p
:
/
/
r
e
s
t
u
d
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

rdq028 2011/6/14 7:05 page 1056 #15
as age goes to zero one can verify that T
k,k
(a) 1 as age goes to zero. The solution for T
0,k
(a)
follows directly from T
1,k
(a) and integrating equation (26). Summing T
n,k
(a) over all n N
gives 1T
1,k
(0)T
0,k
(0) =e
a
(1[ (b)/]
k
) and then T
1,k
(0) follows from integrating
equation (25). The proof of Proposition 5 can be completed by computing the derivative of
T
n,k
(a) and checking equation (27) for any n N. Appendix C gives a more constructive proof
based on the observation that, conditional on no transition, a rm with n blueprints gains and
loses blueprints with the same probabilities as does the aggregate of n independent rms with
one blueprint each.
If =0, then T
0,k
(a) min{1, /} as the age of a cohort grows without bound. If < ,
then virtually all of a cohort of rms will have exited the economy after a sufciently long time.
On the other hand, if > , then a fraction 1 / of any cohort of rms survives and grows
forever, giving rise to a thick-tailed size distribution.
3.4.2. Age given size. Now consider the setup of Proposition 4, with a fraction of a
cohort of new rms entering with high-quality blueprints. Write T
H,n,k
(a) and T
L,n,k
(a) for the
solutions to equations (25)(27) associated with the parameters (
H
,
H
,
H
) and (
L
,
L
,
L
),
respectively. Let n =1 now represent the absorbing state low-quality rms enter into at a rate
L
. Then, the cohort size distribution {p
n
(a)}
n=1
at age a is given by
p
n
(a) =(1)T
L,n,1
(a) +
_
T
H,n,1
(a) +
H
_
a
0
_

k=1
T
L,n,k
(b)T
H,k,1
(a b)
_
db
_
(28)
for all n +1 N. For n = 1, the term T
H,n,1
(a) drops out since high-quality rms do not
transition directly into the state n =1. The innite sum on the right-hand side of equation (28)
can be calculated explicitly, as reported in Appendix C. The rst two terms on the right-hand
side of equation (28) account for the rms that are still in their original growth phase. A ow
H
T
H,k,1
(a b) of high-quality rms with k blueprints transition to become low-quality rms at
age a b. Adding up over all sizes and ages and accounting for subsequent rm growth gives
the third term. Note well that only a fraction 1 p
1
(a) p
0
(a) of the cohort survives until age
a as active rms.
Along a balanced growth path, the measure of entering rms is growing at a rate . Consider
the population of all rms that have ever entered up to a particular point in time, including
those that have since exited. The exponential rate at which the size of entering cohorts grows
implies that this population has an exponential age distribution with density e
a
. Because
{p
n
(a)}
n=1
includes rms that have exited, the joint density of age and size is e
a
p
n
(a)
among all rms that have ever entered. The age density among all rms of size at least n is
therefore
h
n
(a) =
e
a
k=n
p
k
(a)
_
0
e
b
k=n
p
k
(b)db
. (29)
In particular, for n =1, this denes the age density among all surviving rms.
3.5. Gibrat and Zipf: A convenient limiting case
Consider again the economy in which all rms have the same growth parameters and . As
noted after Proposition 1, if the entrepreneurial talent distribution has bounded support, then the
economy may have a balanced growth path with =. Along such a balanced growth path,
there is no entry by new rms, incumbent rms grow randomly, and there is no stationary size
distribution.

a
t

Q
u
e
e
n
'
s

U
n
i
v
e
r
s
i
t
y

o
n

J
u
l
y

8
,

2
0
1
4
h
t
t
p
:
/
/
r
e
s
t
u
d
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

rdq028 2011/6/14 7:05 page 1057 #16
Alternatively, suppose the entrepreneurial talent distribution does have unbounded support,
implying > and the existence of a stationary size distribution along the balanced growth
path. Consider a sequence of such economies for which approaches frombelow. Suppose
that is positive and bounded away from zero, so that the tail index =/() converges to
1. Suppose further that is also positive and bounded away from zero along the sequence. Then,
the recursion (21) for Q
n
nP
n
converges to a limiting recursion that can be written as P
n
=
(P
n+1
+X
n+1
), where X
n+1
=
_
n1
n+1
_
X
n
for all n 1 N. Note that X
n+1
=2X
2
/[n(n +1)]
for all n N. Iterating forward on the recursion for P
n
and requiring the resulting P
n
to add up
to 1 then yields
P
n
=
1
ln(/)
k=n
(/)
k+1n
k(k +1)
.
Thus, the sequence of stationary size distributions has a well-dened limiting distribution. The
right-tail probabilities of this limiting distribution satisfy
lim
n
n
k=n
P
k
= lim
n
1
ln(/)
m=0
n
n +m
_
_
m+1
=
/ 1
ln(/)
, (30)
by the dominated convergence theorem. Thus, the right-tail probabilities behave like 1/nthey
satisfy Zipfs law.
This limiting size distribution implies a limiting rm entry rate = +P
1
> > 0, even
though the rate =() at which new blueprints are introduced by new rms converges to
zero. The average rmsize / grows without bound, as expected fromequation (24). Evaluating
P
1
and using = gives
=
/ 1
ln(/)
. (31)
When the average incumbent rm is very large, virtually all blueprints are created by incumbent
rms, even though the rm entry rate is strictly positive. Under such circumstances, equation
(31) allows one to infer and simply from the population growth rate and the rm entry
rate .
The formulas (30) and (31) provide a simple way to infer the number of employees per
blueprint. A comparison of equations (30) and (31) shows that / nR
n
for large n. The higher
the entry rate, the more rms there will be in the right tail. If R
n
is measured by the fraction of all
rms with more than a certain large number of employees, then (/)/R
n
provides an estimate
of the associated number of blueprints n, and this then implies an estimate for the number of
employees assigned to each blueprint.
Given only one growth phase, the expression (29) for the density of rm age given a size
of at least n blueprints reduces to h
n
(a) e
a
k=n
T
k,1
(a) =e
a
[1(/) (a)]
n1
(a),
where (a) is given in Proposition 5. In the Zipf limit = , this depends on a only via
e
a
. Setting the derivative of h
n
(a) equal to zero then shows that the mode a
mode
of this density
satises
a
mode
=ln
_
1+
n 1
/
_
(32)
for all n N. Letting 0 so that gives the Zipf asymptote of the Yule process, and this
yields a
mode
= ln(n)/. In this limit, the mode of the age density among rms of size no less
than n turns out to be equal to the time it would take a rm to grow from 1 to n at a deterministic
rate = . Increasing subject to = increases the variability of rm growth rates and
spreads out the size distribution of an age cohort. This makes it more likely that surviving rms
of a given age are large, which lowers the modal age of large rms. In fact, formula (32) shows

a
t

Q
u
e
e
n
'
s

U
n
i
v
e
r
s
i
t
y

o
n

J
u
l
y

8
,

2
0
1
4
h
t
t
p
:
/
/
r
e
s
t
u
d
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

rdq028 2011/6/14 7:05 page 1058 #17
that the modal age of large rms converges to zero as / increases without bound. But equation
(31) implies that then the rm entry rate also grows without bound. This will be the tension that
makes it hard to hold on to Gibrats law.
4. U.S. EMPLOYER FIRMS
U.S. Internal Revenue Service statistics contain more than 26 million corporations, partnerships
and non-farm proprietorships. Business statistics collected by the U.S. Census consist of both
non-employer rms and employer rms. In 2002, there were more than 17 million non-employer
rms, many with very small receipts, and close to 6 million employer rms.
Here, Census data on employer rms assembled by the U.S. Small Business Administration
(SBA) will be considered. For employer rms, part-time employees are included in employee
counts, as are executives. But proprietors and partners of unincorporated businesses are not
(Armington, 1998, p. 9). This is likely to create signicant distortions in measured employment
for small rms. The SBA reports rm counts for 24 size categories, ranging from 1 to 4 em-
ployees to 10000 and more employees, as well as the number of employer rms that have no
employment in March but some employment at other times during the year. Over the period
19892006, SBA data show that the number of rms grows roughly at the population growth
rate of about 1% per annum, in line with the theory presented here and in Luttmer (2007).
Newly collected data on the age of rms with more than 10 000 employees in 2008 will
also be used. Two measures of rm age are reported in Figure 3 below. One is based on the
date a rm was incorporated. Corporate restructuring can cause this measure of age to be much
below the age of the underlying organization that constitutes the rm. An alternative measure
uses the earliest date a rm or any of its components are known to have been in operation. A
more detailed description of how this data was collected is given in Appendix A. Clearly, the
complicated genealogy of many large corporations is not captured by the models described in
this paper.
4.1. Gibrat implies 750-year-old rms
Panels (i) and (ii) of Figure 2 show the tted employment size distribution assuming that there
is only one growth phase. The fractions #{rms with employment x}/#{all rms} and #{rms
with employment x}/#{all rms} observed in the data are displayed after merging the category
of employer rms with no employment in March with the category of 1 to 4 employees. The
right tail of the size distribution, shown in panel (ii), is clearly well approximated by x
, and
the slope of the log tail probabilities with respect to x is about 1 05. Note that this estimate
does not depend on the units in which rm size is measured. The formula for the tail index
combined with a 1% population growth rate implies that incumbent rms grow at a rate =
/ 0 95% per annum.
To decompose , consider rst the Yule process obtained by setting = 0 and = 0
0095. The only remaining free parameter is then the number of employees per blueprint i + j +l.
To see how this parameter can be identied, write =i + j +l and recall that R
n
is the fraction
of rms with n or more blueprints. The fraction of rms with at least x {n : n N} employees
is then R
x/
. The vertical axis of panel (ii) of Figure 2 shows ln(R
x/
) and the horizontal axis
shows ln(x) =ln() +ln(x/). Thus, an increase in employment per blueprint moves the model
prediction [ln(x), ln(R
x/
)] to the right by the change in ln(), for every ln(R
x/
). Since the
empirical counterparts to [ln(x), ln(R
x/
)] are pretty much on a straight line for all rms with
more than 10 employees, the choice of will simultaneously either t or fail to t all right tail
probabilities for rms with more than 10 employees. The close t of the right tail shown in Panel

a
t

Q
u
e
e
n
'
s

U
n
i
v
e
r
s
i
t
y

o
n

J
u
l
y

8
,

2
0
1
4
h
t
t
p
:
/
/
r
e
s
t
u
d
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

rdq028 2011/6/14 7:05 page 1059 #18
(ii) is obtained by setting i + j +l =2. Panel (i) shows that the left tail is also well approximated.
The stationary size distribution of a Yule process ts the empirical size distribution quite well.
But a Yule process predicts no exit and a rm entry rate equal to = or about 1% per annum.
Instead, the SBA reports a rm entry rate of about 11% per annum over the period 19892006.
Actual rms do decline and exit, and entry rates are much higher than the population growth
rate.
To match the evidence on rm entry, consider rst the 1 approximation. Given =0 01
and =0 11, solving equation (31) for and = gives =0 4216 and =0 4116. In
2008, close to 1000 rms out of a total of 6 million employer rms had at least 10 000 employees.
The combination of equations (30) and (31) therefore gives i + j +l 10 000R
n
/(/)
10 (1/6)/(0 11/0 01) = 0 15. The more precise estimates obtained for = 1 05 are very
similar. Increasing and subject to the constraint = / 0 0095 until the implied
entry rate reaches the 0 11 value observed in the data yields = 0 4095 and = 0 4000.
Choosing the number of employees per blueprint to match the right tail probabilities now implies
i + j +l = 0 20. The associated left and right tails are shown in Panels (i) and (ii) of Figure
2. The increased transition probabilities and raise the variance (+)/n of the growth
rate of a rm with n blueprints, and this implies that surviving rms are more likely to have
many blueprints. Fitting the right tail of the employment distribution therefore requires fewer
employees per blueprint than in the case of a Yule process. But then the left tail of the size
distribution no longer ts well. The higher variance cuts down, too much, on the number of
small rmsthey either exit or grow large.
The age distributions displayed in the upper panel of Figure 3 show a much more dramatic
failure of the one-phase model of rm growth. At =0 4095, =0 4000 and i + j +l =0 20,
the median age of rms with more than 10 000 employees is about 750 years. The Yule process
tted above implies a median large rm that is a couple of centuries older still. In the data, the
median age of these large rms is only about 75 years. Given Gibrats law, all rms grow at the
same average rate , and this cannot exceed , or about 1% per year. Deterministic growth
would imply that it takes ln(50000)/0 01 1082 years to reach the size of 10 000 employees.
The 1 approximation (31) for the entry rate gives / 42 16 and then the approximation
(32) for the mode of the age distribution of large rms gives a
mode
= 100 ln(1 +49999/42
16) 708 years. Adding variability lowers the age of the typical large rm, but the amount of
variability that can be added is constrained by the entry and exit evidence.
SBA data for the period 19892006 show that the annual exit rate of rms with 500 or more
employees is about 25%. As calibrated so far, the model implies that this number should be
essentially zero. The annual growth rate of a rm with 500 employees has a standard deviation
of only
(0 4095+0 4000)/2500 0 018 and rms only exit after losing all blueprints. To
account for the observed exit of large rms, suppose a rm may not only exit after losing its last
blueprint but also randomly at a rate = 0 02. The approximation 1 then implies that sur-
viving rms grow at a rate =+ of about 3% per year. The entry approximation (31) ob-
tained by replacing with + yields /(+) 9 1, and hence 0.2730 and 0.2430.
With a tail index =1.05 instead of =1, this becomes =0.2700 and =0.2415. Given that
random exit now accounts for 2% of the 10% exit rate, randomness at the blueprint level must
be lower than before. As a result, there will be more small rms, and adjusting the number of
employees per blueprint to match the right tail of the size distribution now gives i + j +l 0.6.
Thus, a rm with 10 000 employees has about 16 667 blueprints, and the approximation (32)
then implies a
mode
= ln(1 +16666/9.16)/0.03 250, a drastic improvement over the cali-
bration without random exit. The entire age distribution is shown in Figure 3. Deterministic
growth conditional on survival would give ln(50000)/0.03 361 or ln(16667)/0.03 324,
and so much of the reduction in age is simply due to the fact that survivors can now grow at

a
t

Q
u
e
e
n
'
s

U
n
i
v
e
r
s
i
t
y

o
n

J
u
l
y

8
,

2
0
1
4
h
t
t
p
:
/
/
r
e
s
t
u
d
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

rdq028 2011/6/14 7:05 page 1060 #19
FIGURE 2
Left and right cumulative distribution functions
a 3% annual rate instead of a 1% annual rate. But this calibration still predicts that the median
U.S. rm is older than the U.S. itself.
4.2. Rapid initial growth
As Figure 1 suggests, many large rms became large during relatively short periods of growth
at rates far exceeding the sum of the population growth rate and the exit rate of large rms. This
can account for the fact that the median large rm is only 75 years old. Proposition 4 indicates
how this can also be made consistent with the observed right tail of the size distribution. Firms
can grow initially at a high rate
H
H
and then transition at a rate
H
to a regime with a
growth rate
L
L
that must be below +
L
. If the tail index is determined by the effects of
initial rapid growth, then = ( +
H
)/(
H
H
). Given 1.05 and 0.01, this implies
that
H
H
must be close to
H
. An initial phase with very rapid growth is possible as long
as this phase is of sufciently short average duration. The realized durations of the high-growth
regime are exponentially distributed, implying that some rms grow rapidly for much longer
than the average duration. This results in relatively young large rms.
Allowing for an initial growth phase adds the parameters ,
H
,
H
,
H
and i
H
+ j
H
+l
H
. This
gives more than enough exibility to match the observed median age of large rms. The theory
of Section 2 implies i
H
+ j
H
+l
H
> i
L
+ j
L
+l
L
but is silent on the magnitude of the difference.
Measured employment per blueprint may not reect effective labour used per blueprint if work-
ers differ in ability. Panels (iii) and (iv) of Figure 2 and the lower panel of Figure 3 show the size

a
t

Q
u
e
e
n
'
s

U
n
i
v
e
r
s
i
t
y

o
n

J
u
l
y

8
,

2
0
1
4
h
t
t
p
:
/
/
r
e
s
t
u
d
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

rdq028 2011/6/14 7:05 page 1061 #20
FIGURE 3
Distributions of rm startup years
and age distributions for a benchmark calibration in which a single employee is assigned to ev-
ery blueprint. A fraction =0.4 of new rms start out as high-growth rms. In the low-growth
regime, the rates at which rms gain and lose blueprints are
L
=
L
=0.2500, and rms in this
regime exit randomly at a rate
L
=0.02.
13
In the high-growth regime, rms also lose blueprints
at the rate
H
=0.2500, but they gain blueprints at the higher rate
H
=0.3825. Firms transition
from the high-growth to the low-growth regime at a rate
H
= 0.125, resulting in a tail index
= ( +
H
)/(
H
H
) 1.02. These parameters imply an exit rate of about 11% and Fig-
ures 2 and 3 show that these parameters closely match the observed size and age distributions.
Holding xed the other parameters, increasing the fraction of high-growth new rms lowers
the entry rate below and increases the number of large rms above what is observed in the data.
Jointly changing (
H
,
H
) to (
H
+,
H
+) changes the exit rate and the size distribution
very little but shifts the age distribution of large rms. At
H
=
L
= 0.02, the age distribution
is essentially that of the one-regime economy, while the median age of large rms can be as low
as 50 years when
H
=0.25.
But what about Gibrats law? Many researchers nd that Gibrats law is a good approxima-
tion for rms that are not too small (e.g. Hall, 1987; Evans, 1987). Figure 4 shows the mean and
standard deviation of the growth rates of surviving rms conditional on employment size. The
mean growth rates are shown for three horizons: instantaneous, 1 year and 5 years. Survivorship
bias affects the instantaneous mean only for rms with one employee. The probability of the
13. The U.S. Bureau of Labor Statistics reports monthly job separation rates for the 2000s that add up to around
50% on an annual basis. The parameter values
H
=
L
=0.25 can be interpreted to mean that half of these separations
do not correspond to job destruction, but to worker replacement.

a
t

Q
u
e
e
n
'
s

U
n
i
v
e
r
s
i
t
y

o
n

J
u
l
y

8
,

2
0
1
4
h
t
t
p
:
/
/
r
e
s
t
u
d
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

rdq028 2011/6/14 7:05 page 1062 #21
FIGURE 4
Growth rates by rm size
multiple blueprint losses it takes to induce exit is second order for rms with more than one
employee. For the longer horizons commonly used in empirical studies this is no longer true,
and surviving small rms grow much faster on average than the unconditional mean. But this
effect declines rather quickly with size, and mean growth rates do not show much variation with
size overall (the horizontal scale is logarithmic).
Over short intervals of time, the variance of surviving rm growth rates in phase I {H, L}
is (
I
+
I
)/n for rm with n > 1 blueprints and
I
for rms with one blueprint. The calibration
implies that this adds up to a standard deviation among all surviving rms of about 41% per
annum. This is well within the range of standard deviations reported in Davis et al. (2007). For
rms with more than one blueprint, the calibration implies a standard deviation of a rm with
n blueprints is about 0.80/
n in the high-growth phase and 0.71/
n in the low-growth phase.

This makes for very volatile growth rates among small rms. But for rms with more than 10 000
employees, these standard deviations will be less than 1% per annum. Even for a model without
aggregate shocks, this is probably too small.
As emphasized by Klette and Kortum (2004), the empirical evidence suggests that the vari-
ance of rm growth rates declines more slowly than 1/n. Hymer and Pashigian (1962) compared
standard deviations of rm growth rates across size quartiles and found that rms in the largest
quartile were signicantly more volatile than predicted by the 1/n rule. More recently, Stanley
et al. (1996) and Sutton (2002) nd that the variance of the growth rate of Compustat rms
behaves like 1/n
1/3
, and tentative interpretations are given in Stanley et al. (1996) and Sutton
(2002, 2007). However, most rms are not publicly traded and are not covered by Compustat,
and it is possible that these studies miss a rapid decline in variance that happens for small n.

a
t

Q
u
e
e
n
'
s

U
n
i
v
e
r
s
i
t
y

o
n

J
u
l
y

8
,

2
0
1
4
h
t
t
p
:
/
/
r
e
s
t
u
d
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

rdq028 2011/6/14 7:05 page 1063 #22
5. CONCLUSION
Skewed rm size distributions are interpreted as reecting skewed productivity distributions in
Hopenhayn (1992), Atkeson and Kehoe (2005) and Luttmer (2007), among many others. The
current paper attributes size differences not only to productivity differences but also to stochas-
tic variation in the number of markets in which a rm operates, as in Klette and Kortum (2004),
Lentz and Mortensen (2006) and Arkolakis (2006). Bounded productivity differences may give
rise to unbounded size differences. In Lucas (1978), all variation in rm size is determined by
heterogeneity in managerial talent. In Holmes and Schmitz (1995), Gabaix and Landier (2008)
and Tervi (2008), both rm-specic productivity and managerial productivity play a role. Much
remains to be done to sort out the relative importance of each of these aspects of rm hetero-
geneity.
Figure 1 and the relative young age of large rms are interpreted here using a two-phase
pattern of growth in which some new rms start out with a high-quality blueprint and become
rms with all low-quality blueprints after some random time. This is an abstraction that helps
to illustrate the type of growth mechanism that can explain the size and age distribution of large
rms. One expects more gradual declines in relative quality to work as well. A natural extension
would allow for start-up blueprints that are initially of uncertain quality. This would bring in the
selection considerations emphasized by Jovanovic (1982).
If blueprints are location specic, and locations are known to differ in how protable they
can be, then rms with new ideas will initially implement these in the more protable locations,
and only then expand, at a slower pace, into less attractive locations. This could be an alternative
interpretation of the growth patterns shown in Figure 1, although it remains to be seen how this
can account for the observed size distribution. One possibility is suggested by static models of
Pareto-like size distributions. A well-known example is the Beckmann (1958) model of hierar-
chies of cities. More recently, Hsu (2007) describes an equilibrium model of hierarchies of rms
and cities that produces Zipfs law. These static models could be viewed as long-run equilibrium
conditions for a dynamic economy, and then the rapid initial growth shown in Figure 1 would
simply reect the fact that setting up a large rm is not quite instantaneous but still very fast.
Firms can grow along many margins. They can introduce new goods, build new plants, open
new sales ofces, hire new workers, win new customers, acquire whole new divisions. Entry
sizes and the increments by which rms grow can vary across markets and industries. The
framework sketched in this paper can be extended to incorporate these elements and arrive at
a richer description of rm growth and heterogeneity. A close examination of the early histories
of large U.S. corporations, such as those shown in Figure 1 and the ones described in Appendix
A, shows that mergers, acquisitions and spin-offs are by no means infrequent. Along the lines of
Jovanovic and Rousseau (2002), it is possible to interpret a small acquisition as the production
of a new blueprint, but other interpretations are perhaps more natural. Spin-offs can give rise to
rms that enter with a relatively large initial size, instead of the common minimum size assumed
in this paper. It would be interesting to know if an account can be given of these aspects of rm
growth that is consistent with the observed age and size distributions.
APPENDIX A
A.1. Firm employment and age data
The employment histories shown in Figure 1 were collected from Compustat, historical Moodys Manuals and corpo-
rate Web sites. Most employment histories are incomplete. Abbott Laboratories, Dow Chemical, IBM and Procter and
Gamble were founded in the 19th century, and all other companies shown in Figure 1 during the 20th century.
The rm age data used in Section 4 were collected from several sources. Large rms are taken to be all Compustat
rms headquartered in the U.S. with more than 10 000 employees, and rms in the same size category that appear on

a
t

Q
u
e
e
n
'
s

U
n
i
v
e
r
s
i
t
y

o
n

J
u
l
y

8
,

2
0
1
4
h
t
t
p
:
/
/
r
e
s
t
u
d
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

rdq028 2011/6/14 7:05 page 1064 #23
a list of large privately held U.S. companies published by Forbes magazine, both in 2008. Compustat and Forbes tend
to use a broad measure of rm employment. Employment at foreign subsidiaries is included, and franchisee employees
appear to be counted as employees of the franchisor. The number of large rms obtained in this way is 813, which is
somewhat less than the 953 rms with more than 10 000 employees reported by the SBA for 2006. In part, this may be
because rms headquartered outside the U.S. are not included here. For example, Shell Oil Company is incorporated in
the U.S. but is a wholly owned subsidiary of an English company headquartered in the Netherlands.
The date of incorporation of publicly traded rms was taken from the synopsis section of the Mergent Online
database. For all rms, the foundation date is the date of the earliest reference to the company or its known predecessor
companies that can be found in the company history section of the Mergent Online database or on company Web sites.
In cases where this information is not available or does not appear to refer to the earliest times of the company, three
additional sources were consulted: Dun and Bradstreets Million Dollar Database, Hoovers Company Reports and the
International Directory of Company Histories. In the case of privately held companies, Hoovers is the primary source.
In a few cases, company age data were found in the Encyclopedia Britannica or in books available in the Google Books
online library.
The Mergent Online database contains extensive records on now defunct corporations that were sometimes used to
further trace back the origins of a company. Corporate Web sites of large companies often include extensive company
histories that tend to emphasize the very old roots of the rm. Occasionally, the foundation date is taken to be the date
its founding entrepreneur rst started a business in the same industry, even if the company that eventually became large
was not the rst company started by the entrepreneur.
In the data collected here, mergers are an important source of rm growth that is not accounted for by the models
in this paper. In many cases, a company history includes the year in which the oldest known component of a rm was
founded. But in some industries, notably the health care industry, such information is almost non-existent. The models
in this paper also do not allow for spinoffs. In the data set constructed here, a company that was already large at the time
of its spinoff is taken to be founded at the time its parent was founded, when this information is available. Clearly, future
empirical and theoretical work needs to account explicitly for mergers and spinoffs. The company age data together with
the source for each age observation are available at www.luttmer.org.
APPENDIX B
B.1. Proof of Propositions 2, 3 and 4
Throughout this appendix, let , = and () be positive, and take {A
n
}
n=1
to be non-negative, not identically
zero, and summable. Let Y
0
=0 and consider the difference equation
1
n
Y
n
=Y
n+1
+Y
n1
(+)Y
n
+
A
n
n
(B.1)
for all n N. Dene
0
=0 and
1
n
=1+
+n
n

n1
(B.2)
for all n N. Then, the second-order difference equation (B.1) is equivalent to the pair of rst-order difference equations
Y
n
= Z
n
+
n
Y
n+1
, Z
n
=

n
_
Z
n1
+
A
n
n
_
(B.3)
for all n N, combined with the initial condition Z
0
=0.
Lemma A1. The sequence {
n
}
n=0
increases monotonically from 0 to min{1, /}. If > , then lim
n
n(1
n
/) =/(). If < , then lim
n
n(1
n
) =/().
The proof can be given using diagrams for the recursions that dene
n
, n(1
n1
/) and n(1
n1
), respectively.
Note that n =in equation (B.2) yields a quadratic equation that is solved by
{1, /}.
It is convenient to dene two integrating factors,
B
n
=
1
n
n
k=1
k
, C
n1
=
_
_
n1
B
n
for all n N. Note that B
1
= C
0
= 1, C
n
/C
n1
=
n
/, and equation (B.3) can be written as B
n
Y
n
= B
n
Z
n
+
B
n+1
Y
n+1
and Z
n
/C
n
= Z
n1
/C
n1
+ A
n
/(nC
n1
). We are interested only in solutions to equation (B.1) that are

a
t

Q
u
e
e
n
'
s

U
n
i
v
e
r
s
i
t
y

o
n

J
u
l
y

8
,

2
0
1
4
h
t
t
p
:
/
/
r
e
s
t
u
d
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

rdq028 2011/6/14 7:05 page 1065 #24
non-negative and summable. Since B
n
1, this implies that any such solution must satisfy lim
n
B
n+1
Y
n+1
= 0.
Given this boundary condition, equation (B.3) implies that
Y
n
=
1
B
n
k=0
B
n+k
C
n+k
n+k
j =1
A
j
j C
j 1
(B.4)
for all n N. The following lemma collects some facts that are useful in determining the properties of Y
n
.
Lemma A2. Let {a
n
}
n=1
be a positive sequence. If lim
n
a
n+1
/a
n
= (0, 1), then
lim
n
na
n
n
k=1
1
ka
k
= lim
n
1
a
n
k=0
a
n+k
=
1
1
. (B.5)
Alternatively, if lim
n
n(1a
n+1
/a
n
) = (0, ), then
lim
n
a
n
n
k=1
1
ka
k
= lim
n
1
a
n
k=0
a
n+k
n +k
=
1
(B.6)
and Raabes test says that {a
n
}
n=1
is summable if > 1 and not if < 1.
A version of equation (B.6) for regularly varying functions is discussed as part of Karamatas theorem in Bingham,
Goldie and Teugels (1987). See Bojanic and Seneta (1973) for regularly varying sequences.
Suppose now that all A
n
are positive and that lim
n
n(1A
n+1
/A
n
) =
A
(0, ] is well dened. Also write
= lim
n
A
n+1
/A
n
in situations in which this limit is well dened. Dene
C
= lim
n
n(1 C
n+1
/C
n
). By
Lemma A1,
C
=/() > 1 if > and
C
=if < . The large-n behaviour of Y
n
can be inferred from the
fact that equation (B.4) can be represented as follows:
Y
n
= T
n
k=0
w
n,k
x
n+k
, (B.7)
where T
n
equals either A
n
or C
n
, and
lim
n
k=0
w
n,k
=, lim
n
x
n
= (B.8)
for some and in [0, ). This implies that lim
n
Y
n
/T
n
=.
When C
n
declines more slowly than A
n
, the following representation works
T
n
=C
n
, w
n,k
=
B
n+k
C
n+k
B
n
C
n
, x
n
=
n
j =1
A
j
j C
j 1
. (B.9)
As long as = , Lemma A1 implies B
n+1
C
n+1
/(B
n
C
n
) min{/, /} (0, 1) and then equation (B.5) gives
=1/(1min{/, /}). If / > 1 and
C
<
A
, then Raabes test implies that x
n
converges to some <.
The same is true if < / < 1.
When > and A
n
declines more slowly than C
n
, consider the representation
T
n
= A
n
, w
n,k
=
A
n+k
B
n+k
A
n
B
n
, x
n
=
_
A
n
C
n
_
1
n
j =1
A
j
j C
j 1
.
NowLemma A1 implies A
n+1
B
n+1
/(A
n
B
n
) /(0, 1) and equation (B.5) gives =1/(1/). Since
A
<
C
,
(B.6) implies =1/(
C
A
).
When < and A
n
declines more slowly than C
n
, consider
T
n
= A
n
, w
n,k
=
1
n +k
A
n+k
B
n+k
A
n
B
n
, x
n
=
nC
n
A
n
n
j =1
A
j
j C
j 1
.
Lemma A1 implies n(1B
n+1
/B
n
) /() > 0 and C
n
/C
n1
/ (0, 1). Furthermore, n(1A
n+1
B
n+1
/
(A
n
B
n
))
A
+/( ) > 0 and thus (B.6) implies = 1/[
A
+/( )]. If
A
< then (C
n+1
/A
n+1
)/
(C
n
/A
n
) / (0, 1). Alternatively, if
A
=and / < <1, then (C
n+1
/A
n+1
)/(C
n
/A
n
) [/]/ (0, 1).
In either case, equation (B.5) implies that x
n
converges to some < .
Finally, note that the representation (B.7)(B.9) also works and lim
n
Y
n
/C
n
(0, ) if A
n
>0 for only nitely
many n.

a
t

Q
u
e
e
n
'
s

U
n
i
v
e
r
s
i
t
y

o
n

J
u
l
y

8
,

2
0
1
4
h
t
t
p
:
/
/
r
e
s
t
u
d
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

rdq028 2011/6/14 7:05 page 1066 #25
Proposition A1. Consider the cases (i) A
n
> 0 for nitely many n, (ii) lim
n
A
n+1
/A
n
= and (iii)
lim
n
n(1 A
n+1
/A
n
) =
A
(0, ]. Then, lim
n
Y
n
/C
n
(0, ) in case (i), in case (ii) if 1 > / > and
in case (iii) if
A
>
C
. Alternatively, lim
n
Y
n
/A
n
=0 in case (ii) if 1 > > /, and lim
n
Y
n
/A
n
(0, ) in
case (iii) if
A
<
C
.
If we add the restriction
A
> 1 when A
n
> 0 for all n, then Raabes test implies that the solution (B.4) is in fact
summable in all cases. One can generate many more solutions by adding a positive constant to B
n
Y
n
, but these cannot
be summable.
Proposition A2. Suppose A
n
> 0 for nitely many n, or lim
n
n(1 A
n+1
/A
n
) =
A
> 1. Then,
n=1
Y
n
=
1
()
n=1
A
n
,
n=1
1
n
Y
n
=
1
_
_
n=1
A
n
n
Y
1
_
_
.
Proof. It follows from min{
A
,
C
} > 1 that nY
n
0. Recall Y
0
=0 and write equation (B.1) as follows:
Y
n
=[(n +1)Y
n+1
nY
n
Y
n+1
] [nY
n
(n 1)Y
n1
Y
n1
] +A
n
for all n N. Adding up over all n and using lim
n
nY
n
=0 gives the rst sum. The second sum follows from directly
summing equation (B.1).
With these results in place, Proposition 2 follows from taking = , A
1
= ( ())/ and A
n+1
= 0 for n N.
Proposition 3 then follows since part (B.6) of Lemma A2 shows that

k=n
Y
k
/k behaves like Y
n
C
n
, and
C

(1, ).
To prove Proposition 4, let Q
T,n
be the fraction of all blueprints in the economy held by type-T rms with n
blueprints, where T {H,L} and n N. The properties of Q
H,n
are an application of Propositions 2 and 3. Proposition
4 can be shown by taking = +
L
, (, ) =(
L
,
L
), A
1
=[(1) +
H
Q
H,1
]/
L
, and A
n+1
=
H
Q
H,n+1
/
L
for all n N.
APPENDIX C
C.1. Proof of Proposition 5
C.1.1. Preliminaries. Suppose {X
i
, Y
i
}
k
i =1
are 2k independent random variables with Pr[X
i
= n] =
(1 )
n1
, n N, Pr[Y
i
=0] =, and Pr[Y
i
=1] =1. Dene Z
k
=
k
i =1
X
i
Y
i
and let K
k
=
k
i =1
Y
i
.
As can be veried using moment generating functions, the sum of i.i.d. geometrically distributed random variables
has a negative binomial distribution, given by
Pr
_
_
m
i =1
X
i
=n
_
_
=
_
n 1
m1
_
(1 )
m
nm
for all m N and n +1m N. In view of the independence assumptions,
Pr[Z
k
=n] =Pr
_
_
k
i =1
X
i
Y
i
=n
_
_
=
min{k,n}
m=1
Pr
_
_
m
i =1
X
i
=n
_
_
Pr[K
k
=m]
for all n N. Using the binomial distribution of K
k
, this implies
Pr[Z
k
=n] =
min{k,n}
m=1
_
k
m
__
n 1
m1
_
(1)
m
km
(1 )
m
nm
(C.1)
for all n N. The complementary probability is Pr[Z
k
= 0] =
k
since Z
k
= 0 if and only if all Y
i
are zero. Also,
n=1
n Pr[Z
k
=n] =k (1)/(1 ). This can be used to compute the mean growth rates conditional on survival
reported in Figure 4.

a
t

Q
u
e
e
n
'
s

U
n
i
v
e
r
s
i
t
y

o
n

J
u
l
y

8
,

2
0
1
4
h
t
t
p
:
/
/
r
e
s
t
u
d
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

rdq028 2011/6/14 7:05 page 1067 #26
Now suppose that K is drawn from the geometric distribution (1)
k1
, k N. Then, the distribution of Z
K
is
determined by
(1)
k=1
k1
Pr[Z
k
=n] =
(1)
n
(1)
n
m=1
_
n 1
m1
__
(1)(1 )
(1)
_
m
(C.2)
for all n N. The right tail probabilities of this distribution are
(1)
n=N
k=1
k1
Pr[Z
k
=n] =
1
1
_
(1) +(1)
1
_
N1
(C.3)
for all N N. For N =1 this yields Pr[Z
k
=0] =(1)/(1).
Sketch of proof and computation. Suppose =0. Consider a rm that starts out with one blueprint.
As reported in Klette and Kortum (2004), by age a such a rm will have exited with probability T
0,1
(a) =

(a). Con-
ditional on survival, its size distribution is the geometric size distribution T
n,1
(a)/[1T
0,1
(a)] = [1 (a)]
n1
(a).
This can be veried directly by checking equations (26)(27). The size distribution at age a of a rm that starts out with
k blueprints is simply the distribution of the aggregate of k independent rms that start with one blueprint. Applying
equation (C.1) gives {T
n,k
(a)}
n=1
for the case =0. Now suppose > 0. Transitions from the rst to the second phase
occur at a rate , as long as no exit has taken place. This means that only a fraction e
a
of surviving rms remain in
the initial phase. This determines {T
n,k
(a)}
n=1
. The formulas for T
1,k
(a) and T
0,k
(a) then follow from integrating
equations (25)(26), as described in the text.
The innite sums needed in equations (28) and (29) follow from equations (C.2) and (C.3). Age densities (distribu-
tions) can then be computed using a univariate (bivariate) numerical integration.
Acknowledgment. This paper has evolved from my Federal Reserve Bank of Minneapolis Working Papers 645,
649 and 657. The data and more detailed proofs are available at www.luttmer.org. I thank Nathalie Pouokam for
skillful research assistance and two anonymous referees for helpful comments. The views expressed herein are those of
the author and not necessarily those of the Federal Reserve Bank of Minneapolis or the Federal Reserve System.
REFERENCES
ARKOLAKIS, C. (2006), Market Access Costs and the New Consumers Margin in International Trade (Working
Paper, University of Minnesota).
ARMINGTON, C. (1998), Statistics on U.S. BusinessesMicrodata and Tables (US Small Business Administration).
ATKESON, A. and KEHOE. P. (2005), Modelling and Measuring Organization Capital, Journal of Political Economy,
113, 10261051.
AXTELL, R. L. (2001), Zipf Distribution of U.S. Firm Sizes, Science, CCXCIII, 18181820.
BECKMANN, M. J., (1958), City Hierarchies and the Distribution of City Size, Economic Development and Cultural
Change, 6, 243248.
BERNARD, A. B., REDDING, S. J. and SCHOTT, P. K. (2006), Multi-Product Firms and Product Switching (NBER
Working Paper No. 12293).
BINGHAM, N. H., GOLDIE, C. M. and TEUGELS, J. L. (1987), Regular Variation (Cambridge: Cambridge University
Press).
BOJANIC, R. and SENETA, E. (1973), A Unied Theory of Regularly Varying Sequences, Mathematische Zeitschrift,
134, 91106.
BOLDRIN, M. and LEVINE, D. K. (1999, 2006), Perfectly Competitive Innovation (Working Paper, Washington
University).
CHATTERJEE, S. and ROSSI-HANSBERG, E. (2006), Spin-offs and the Market for Ideas (Working Paper, Princeton
University).
DAVIS, S. J., HALTIWANGER, J., JARMIN, R. and MIRANDA, J. (2007), Volatility and Dispersion in Business
Growth Rates: Publicly Traded versus Privately Held Firms, in Acemoglu, D., Rogoff, K. and Woodford, M. (eds)
NBER Macroeconomics Annual 2006, Vol. 21 (Cambridge: MIT Press), 107156.
DIXIT, A. K. and STIGLITZ, J. E. (1977), Monopolistic Competition and Optimum Product Diversity, American
Economic Review, LXVII, 297308.
DUNNE, T, ROBERTS, M. J. and SAMUELSON, L. (1989), The Growth and Failure of U.S. Manufacturing Plants,
Quarterly Journal of Economics, CIV, 671698.
EVANS, D. S. (1987), The Relationship between Firm Growth, Size and Age: Estimates for 100 Manufacturing
Industries, Journal of Industrial Economics, XXXV, 567581.
GABAIX, X. (1999), Zipfs Law for Cities: An Explanation, Quarterly Journal of Economics, CXIV, 739367.

a
t

Q
u
e
e
n
'
s

U
n
i
v
e
r
s
i
t
y

o
n

J
u
l
y

8
,

2
0
1
4
h
t
t
p
:
/
/
r
e
s
t
u
d
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

rdq028 2011/6/14 7:05 page 1068 #27
GABAIX, X. and LANDIER, A. (2008), Why Has CEO Pay Increased so Much, Quarterly Journal of Economics,
123, 49100.
GROSSMAN, G. and HELPMAN, E. (1991), Innovation and Growth in the Global Economy (Cambridge: MIT Press).
HALL, B. H., (1987), The Relationship between Firm Size and Firm Growth in the US Manufacturing Sector, Journal
of Industrial Economics, XXXV, 583605.
HOLMES, T. J. (2006), The Diffusion of Wal-Mart and Economies of Density (Working Paper, University of
Minnesota).
HOLMES, T. J. and SCHMITZ, J. A., Jr. (1995), On the Turnover of Business Firms and Business Managers, Journal
of Political Economy, 103, 10051038.
HOPENHAYN, H. (1992), Entry, Exit and Firm Dynamics in Long Run Equilibrium, Econometrica, 60, 11271150.
HSU, W.-T. (2007), Central Place Theory and Zipfs Law (Working Paper, University of Minnesota).
HYMER, S. and PASHIGIAN, P. (1962), Firm Size and Rate of Growth, Journal of Political Economy, 70, 556569.
IJIRI, Y. and SIMON, H. A. (1964), Business Firm Growth and Size, American Economic Review, LIV, 7789.
JACKSON, M. O. (2006), The Economics of Social Networks, in Blundell, R., Newey, W. and Persson, T. (eds)
Proceedings of the 9th World Congress of the Econometric Society (Cambridge: Cambridge University Press) 156.
JACKSON, M. O. and ROGERS, B. W. (2007), Meeting Strangers and Friends of Friends: How Random Are Social
Networks, American Economic Review, 97, 890915.
JOVANOVIC, B. (1982), Selection and the Evolution of Industry, Econometrica, C, 649670.
JOVANOVIC, B. and ROUSSEAU, P. L. (2002), A Q-Theory of Mergers, American Economic Review, Papers and
Proceedings, 92, 198204.
KENDALL, D. G. (1948), On the Generalized Birth-and-Death Process, The Annals of Mathematical Statistics, 19,
115.
KLETTE, T. J. and KORTUM, S. S. (2004), Innovating Firms and Aggregate Innovation, Journal of Political Econ-
omy, 112, 9861018.
KRUGMAN, P. (1996), The Self-Organizing Economy (Oxford: Blackwell Publishers).
LENTZ, R. and MORTENSEN, D. T. (2006), An Empirical Model of Growth Through Product Innovation (University
of Wisconsin and Northwestern University).
LUCAS, R. E., Jr. (1978), On the Size Distribution of Business Firms, Bell Journal of Economics, 9, 508523.
LUTTMER, E. G. J. (2006), Consumer Search and Firm Growth (Working Paper No. 645, Federal Reserve Bank of
Minneapolis.
LUTTMER, E. G. J. (2007), Selection, Growth, and the Size Distribution of Firms, Quarterly Journal of Economics,
122, 11031144.
LUTTMER, E. G. J. (2008), On the Mechanics of Firm Growth (Working Paper No. 657, Federal Reserve Bank of
Minneapolis).
PRESCOTT, E. C. and VISSCHER, M. (1980), Organization Capital, Journal of Political Economy, 88, 446461.
ROMER, P. M. (1990), Endogenous Technical Change, Journal of Political Economy, 98, S71S102.
ROSEN, S. (1978), Substitution and Division of Labour, Economica, 45, 235250.
ROSSI-HANSBERG, E. and WRIGHT, M. L. J. (2007), Establishment Size Dynamics in the Aggregate Economy,
American Economic Review, 97, 16391666.
SEKER, M. (2007), A Structural Model of Establishment and Industry Evolution: Evidence from Chile (Working
Paper, University of Minnesota).
SIMON, H. A. (1955), On a Class of Skew Distribution Functions, Biometrika, 42, 425440.
SIMON, H. A. and BONINI, C. P. (1958), The Size Distribution of Business Firms, American Economic Review,
XCVIII, 607617.
STANLEY, R. M., AMARAL, L. A. N., BULDYREV, S. V., HARLIN, S., LESCHORN, H., MAASS, P., SALINGER,
M. A., STANLEY, H. E. (1996), Scaling Behavior in the Growth of Companies, Nature, 319, 804806.
SUTTON, J. (1997), Gibrats Legacy, Journal of Economic Literature, XXXV, 4059.
SUTTON, J. (2002), The Variance of Firm Growth Rates: The Scaling Puzzle, Physica A, CCCXII, 577590.
SUTTON, J. (2007), Market Share Dynamics and the Persistence of Leadership Debate, American Economic Review,
97, 222241.
TERVI, M. (2008), The Difference that CEOs Make: An Assignment Model Approach, American Economic Review,
98, 642668.
YOUNG, A. (1998), Growth Without Scale Effects, Journal of Political Economy, 106, 4163.
YULE, G. U. (1925), A Mathematical Theory of Evolution Based on the Conclusions of Dr. J.C. Willis, F.R.S.,
Philosophical Transactions of the Royal Society of London, Series B, Containing Papers of a Biological Character,
213, 2187.

a
t

Q
u
e
e
n
'
s

U
n
i
v
e
r
s
i
t
y

o
n

J
u
l
y

8
,

2
0
1
4
h
t
t
p
:
/
/
r
e
s
t
u
d
.
o
x
f
o
r
d
j
o
u
r
n
a
l
s
.
o
r
g
/
D
o
w
n
l
o
a
d
e
d

f
r
o
m

On The Mechanics of Firm Growth: Erzo G. J. Luttmer

Uploaded by

Copyright:

Available Formats

On The Mechanics of Firm Growth: Erzo G. J. Luttmer

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

On The Mechanics of Firm Growth: Erzo G. J. Luttmer

Uploaded by

Copyright:

Available Formats

rdq028 2011/6/14 7:05 page 1042 #1

Review of Economic Studies (2011) 78, 10421068 doi: 10.1093/restud/rdq028

? Why is the tail in-

for any x > 0, where [xn] is the integer part of xn.

n in the high-growth phase and 0.71/

n in the low-growth phase.

You might also like