PhD Topics in Macroeconomics

Lecture 5: innovation and firm dynamics, part one

Chris Edmond

2nd Semester 2014

1
 This lecture

• Review of ‘Schumpeterian’ growth theory

– quality ladders
– endogenous growth via creative destruction

• Appendix: review of some continuous time tools

– continuous time Bellman equations

– Poisson processes

2
 Model overview

• Goods are ‘horizontally’ and ‘vertically’ differentiated

• Vertical differentiation via quality differences

– higher quality varieties deliver more utility per unit consumption

• Quality improvements arrive according to Poisson process

• Arrival rate of quality improvements is endogenous

– gives rise to aggregate growth

3
 Quality ladder model
• Continuous time t 0

• Representative household
Z 1
⇢t
U= e log Ct dt, ⇢>0
0

• Aggregate consumption Ct depends on

– j 2 [0, 1] continuum horizontally differentiated varieties

– k 2 {0, 1, . . . , Jt (j)} discrete vertically differentiated vintages of j

– state-of-the-art vintage Jt (j) for each horizontal variety j

• Let z(j, k) denote quality and xt (j, k) denote quantity of variety j, k

4
 Aggregate consumption

• Instantaneous utility

Z 1 h JX
t (j) i
log Ct = log z(j, k)xt (j, k) dj
0 k=0

• Note: imperfect horizontal differentiation (elasticity of subs. = 1)

but perfect vertical differentiation (elasticity of subs. = 1)

• Let q > 1 denote the size of the quality step, i.e., for each j

z(j, k) = q z(j, k 1) k = 1, 2, . . . , Jt⇤ (j)

• Choose physical units for each variety so that z(j, 0) = 1 for all j.
Then simply z(j, k) = q k for all j

5
 Expenditure
• Let Pt denote aggregate price index associated with Ct and let
Et = Pt Ct denote aggregate expenditure

• Let kt⇤ (j) denote variety that charges lowest price per unit quality

• Demand for variety j, k is then

8
> Et
>
< if k = kt⇤ (j)
pt (j, k)
xt (j, k) =
>
>
:
0 otherwise

• Aggregate expenditure satisfies the intertemporal Euler equation

Ėt Ċt Ṗt

= rt ⇢, , = rt ⇢
Et Ct Pt

6
 Expenditure

• Let Et = 1 be the numeraire. Then from the Euler equation

Ċt Ṗt
rt = ⇢ , =
Ct Pt
• And expenditure on variety j, k is
8
< 1 if k = kt⇤ (j)
pt (j, k)xt (j, k) =
:
0 otherwise

7
 Production

• Wage rate wt per unit labor engaged in production

• Flow profits from production of j, k

⇡t (j, k) = (pt (j, k) wt )xt (j, k)

(i.e., it takes one unit of labor to produce one unit of output, x = l)

• Inelastic aggregate labor supply L. Labor may be employed in

goods production LX or in research LR

LX + LR = L

8
 Pricing

• Consider leader firm with state-of-the-art quality and its closest

follower, one step behind

• Leader has quality advantage q > 1 over follower

• Leader charges limit price pt (j, k) = qwt to prevent entry

• In symmetric equilibrium only the state-of-the-art quality is sold

kt⇤ (j) = Jt (j), and all leaders have flow profits
q 1
⇡t (j, Jt (j)) = (pt (j, Jt (j)) wt )xt (j, Jt (j)) = =: ⇡
q

9
 Innovation

• Any firm can target any product line in an attempt to improve

state-of-the-art

• Technology for innovating: cost c > 0 units of labor time

delivers flow probability dt of successfully innovating to next step
on quality lader, 1 dt flow probability of failure

• Quality steps a Poisson process with intensity >0

• Memoryless process: no advantages of incumbency, potential

entrants can immediately build on the state-of-the-art, etc

10
 Innovation decisions

• Let vt denote the value of an incumbent firm (to be determined)

• ‘Lottery’ delivers vt with flow probability dt or 0 with flow

probability (1 dt), costs wt c dt (in units of labor)

• Expected gain from innovation

vt dt + 0(1 dt) wt c dt

• Gives free-entry condition into innovation

vt  wt c, with equality whenever >0

11
 Incumbents don’t invest in innovation
• Will an incumbent try to get two steps ahead? If successful, charge
price = q 2 w and have sales = 1/q 2 w

• Yields flow profits ⇡ 2 := 1 (1/q 2 ). But leader already has flow

profits ⇡ 1 := 1 1/q even if no investment in innovation

• Incremental profit for investing incumbent

✓ ◆
2 1 1 q 1
⇡ ⇡ =
q q
• Incremental profit for investing entrant
✓ ◆
1 0 q 1
⇡ ⇡ = > ⇡2 ⇡1
q
• With free entry, equilibrium ‘cost of capital’ will be too high for
incumbent firms to find it optimal to invest
12
 Bellman equation for incumbents

• Flow profits ⇡

• Lose incumbency to successful innovator with flow probability dt

• Value vt satisfies continuous time Bellman equation

⇢vt = ⇡ vt + v̇t

(no aggregate risk, all idiosyncratic risk perfectly diversified)

13
 Equilibrium
• We will focus on a stationary equilibrium, constants

(v ⇤ , ⇤
, w⇤ )

• Value for incumbents, from steady-state of Bellman equation

⇡ q 1
v= , ⇡=
⇢+ q
for to be determined

• Labor market clearing

LX + LR = L

with total labor employed in each sector

1
LX = , and LR = c
qw

14
 Free entry condition v  wc
• Case 1: > 0 (innovation). Then v = wc and we can write the
labor market clearing condition
(⇢ + )c
+ c=L
q⇡
or
⇤ L ⇢
=⇡
c q
from which we can then recover v ⇤ = ⇡/(⇢ + ⇤) and w⇤ = v ⇤ /c

• Case 2: = 0 (no innovation). Then LR = 0, LX = L and so

w⇤ = 1/qL, v ⇤ = ⇡/⇢

• Steady-state with innovation exists if quality increments

c
q >1+⇢
L
(i.e., if large population, low discount rate, or low cost etc)
15
 Aggregate consumption

• Recall aggregate consumption index

Z 1 h JX
t (j) i
log Ct = log q k xt (j, k) dj
0 k=0

using z(j, k) = q k for all j

• In equilibrium only latest vintage sold, i.e., xt (j, k) = 1/qwt for

k = Jt (j) and zero otherwise. Hence
Z 1
log Ct = (log q) Jt (j) dj log(q wt )
0

16
 Aggregate growth

• Use LLN to calculate cross-sectional average

Z 1
Jt (j) dj = E[Jt ]
0

where Jt is a Poisson process with intensity ⇤, so

⇤
E[Jt ] = t

• Hence in a steady-state with wt = w ⇤ , aggregate growth is

⇤ Ċt ⇤
g := = (log q)
Ct

17
 Real wages etc

• In steady state, wage w ⇤ is a constant

• But real wage w ⇤ /Pt is growing. Recall that

1 = E t = Pt C t

so

Ṗt Ċt
= = g⇤
Pt Ct
Hence real wage is growing at g ⇤ too

• All growth is due to quality upgrading

18
 Aggregate growth

• If interior equilibrium

⇤ ⇤ ⇤ L ⇢
g = (log q) where =⇡
c q
• So in this case g ⇤ is

– increasing in q (directly, and indirectly via ⇤ )

– increasing in L (scale effect)
– increasing in ⇡ (monopoly profits from successful innovation)

– decreasing in c (barrier to entry/cost of innovation)

– decreasing in ⇢ (greater impatience)

• Otherwise, namely if q < 1 + ⇢c/L, then corner equilibrium with

⇤ = 0 and hence g ⇤ = 0 etc

19
 Next

• Innovation and firm dynamics, part two

• Embedding this in a model of firm dynamics

⇧ Klette and Kortum (2004): Innovating firms and aggregate

innovation, Journal of Political Economy.

• Integrated treatment of firm heterogeneity and growth via creative

destruction

20
 Appendix to Lecture 5

Review of continuous time Bellman equations, Poisson processes

1
 Present and flow values
• Consider the present value
Z 1
⇢(s t)
v(t) = e ⇡(s) ds, t 0
t

• Differentiating with respect to time t

Z 1h i
⇢(s t) ⇢(s t)
v̇(t) = ⇢e ⇡(s) ds + e ⇡(s) ( 1)
t s=t

= ⇢v(t) ⇡(t)

• Commonly written

⇢v(t) = ⇡(t) + v̇(t)

and in steady state, naturally v = ⇡/⇢

2
 Deterministic control
• Now consider the control problem
Z 1
⇢(s t)
v(x(t), t) = max e ⇡(x(s), u(s)) ds, t 0
u(·) t

with state variable x, control u, subject to the law of motion

ẋ(t) = g(x(t), u(t)), x(0) = x0 given

• Value function satisfies the continuous time Bellman equation

h @v i @v
⇢v(x, t) = max ⇡(x, u) + g(x, u) +
u @x @t
where the maximand is the Hamiltonian function

H(x, u, ) = ⇡(x, u) + g(x, u)

with costate variable = @v/@x

3
 Heuristic derivation of the Bellman equation
• Discrete time analogue with period length t > 0, t0 = t + t,
x0 = x + x satisfies
h 1 i
v(x, t) = max ⇡(x, u) t+ v(x0 , t0 )
u 1+⇢ t
• Multiplying both sides by (1 + ⇢ t) and subtracting v(x, t)
h i
⇢v(x, t) t = max ⇡(x, u) (1 + ⇢ t) t + v(x0 , t0 ) v(x, t)
u

• Write out the change in the value function as follows

v(x0 , t0 ) v(x, t) = v(x0 , t0 ) v(x, t0 ) + v(x, t0 ) v(x, t)

v(x0 , t0 ) v(x, t0 ) v(x, t0 ) v(x, t)

= x+ t
x t

4
 Heuristic derivation of the Bellman equation

• Now use the state transition equation x = g(x, u) t to write

0 0 v(x0 , t0 ) v(x, t0 ) v(x, t0 ) v(x, t)

v(x , t ) v(x, t) = g(x, u) t+ t
x t
• Plug change in value function back into Bellman equation
h v(x0 , t0 ) v(x, t0 ) i
⇢v(x, t) t = max ⇡(x, u) (1 + ⇢ t) t + g(x, u) t
u x
v(x, t0 ) v(x, t)
+ t
t
• Divide both sides by t > 0 and take limit as t ! 0 to get
h @v i @v
⇢v(x, t) = max ⇡(x, u) + g(x, u) +
u @x @t

5
 Poisson process
• Continuous time stochastic process x(t) with x(0) = 0 and

(i) stationary independent increments

(ii) increments have Poisson distribution with intensity >0

• Increments x(t0 ) x(t) are independent r.v.’s, for all t, t0

• Increments x = x(t + t) x(t) have Poisson distribution

( t)k e t
Prob[ x = k] =
k!
(depends only on period length t > 0, independent of actual t)

• Sample paths are discontinuous, a ‘counting process’

6
 Properties of Poisson process

• Write the distribution of the increment from date s to s + t

( t)k e t
Prob[ {x(t + s) x(s)} = k] =
k!
that is, a Poisson distribution with parameter t

• In particular, taking s = 0 (and using x(0) = 0) we simply have

( t)k e t
Prob[x(t) = k] =
k!
• So, using standard properties of the Poisson distribution, the
process x(t) has moments

E[x] = Var[x] = t

7
 Properties of Poisson process
• Probability that x = 1 over period of length t is
t
Prob[ x = 1] = ( t)e

• Probability process does not change, x = 0, over same period

t
Prob[ x = 0] = e

(note link to exponential distribution...)

• These can be written

Prob[ x = 1] = t + o( t)
Prob[ x = 0] = 1 t + o( t)

where o( t)/ t ! 0 as t ! 0 and moreover where

Prob[ x > 1] = o( t)

8
 Bellman equation when state is Poisson

• Consider the value function

hZ 1 i
⇢(s t)
v(x(t), t) = E e ⇡(x(s)) ds x(t) , t 0
t

where x(t) follows a Poisson process with intensity

• What is the Bellman equation for this problem?

• Discrete time analogue with period length t > 0, t0 = t + t,

x0 = x + x satisfies
1
v(x, t) = ⇡(x) t + E[v(x0 , t0 ) | x]
1+⇢ t

9
 Bellman equation when state is Poisson
• Multiplying both sides by (1 + ⇢ t) and subtracting v(x, t) gives

⇢v(x, t) t = ⇡(x) t(1 + ⇢ t) + E[v(x0 , t0 ) v(x, t) | x]

• Again write out the change in the value function

v(x0 , t0 ) v(x, t) = v(x0 , t0 ) v(x, t0 ) + v(x, t0 ) v(x, t)

0 0 0 v(x, t0 ) v(x, t)
= v(x , t ) v(x, t ) + t
t
• The latter change is deterministic and can be pulled outside of the
expectation, so

⇢v(x, t) t = ⇡(x) t(1 + ⇢ t) + E[v(x0 , t0 ) v(x, t0 ) | x]

v(x, t0 ) v(x, t)
+ t
t
10
 Bellman equation when state is Poisson
• Now write out the expectation
1
X
E[v(x0 , t0 ) v(x, t0 ) | x] = [v(x + x, t0 ) v(x, t0 )]Prob[ x]
x=0

= [v(x + 0, t0 ) v(x, t0 )]Prob[ x = 0]

+ [v(x + 1, t0 ) v(x, t0 )]Prob[ x = 1]
+ [v(x + 2, t0 ) v(x, t0 )]Prob[ x = 2]
+ ...

• Hence for the Poisson process

E[v(x0 , t0 ) v(x, t0 ) | x] = [v(x+1, t+ t) v(x, t+ t)] t+o( t)

11
 Bellman equation when state is Poisson

• Plug this back into the Bellman equation

⇢v(x, t) t = ⇡(x) t(1 + ⇢ t)

+ [v(x + 1, t + t) v(x, t + t)] t + o( t)
v(x, t0 ) v(x, t)
+ t
t
• Divide both sides by t > 0 and take limit as t ! 0 to get
@v
⇢v(x, t) = ⇡(x) + (v(x + 1, t) v(x, t)) +
@t

12
 Application to quality ladder model
• Let x denote number of quality improvements that have occurred

• Let V (x, t) denote value function of an incumbent firm

• For an incumbent, x matters only if an entrant makes an

innovation — in which case it loses its whole market

• In short, write v(t) := V (x, t) and ⇡(x) = ⇡ so long as x remains

unchanged, write V (x + 1, t) = 0 if entrant innovates, etc

• Then we have the Bellman equation from the Lecture 5 notes

⇢v(t) = ⇡ v(t) + v̇(t)

and in steady-state, naturally v = ⇡/(⇢ + ), i.e., profits

discounted at the risk-adjusted rate ⇢ +

13