Appendix A: Proofs

Proof of Lemma 1. Consider two experts A and B with net worth’s nA B

t and nt , respectively.
Denote by uA B
t and ut the maximal expected utilities that these experts can get in equilibrium
from time t onwards. We need to show that uA A B B A A B B
t /nt = ut /nt . Suppose not, e.g. ut /nt > ut /nt .
Denote by ks , dcs , ϕs ; s ≥ t the optimal dynamic strategy of expert A, which attains utility uA
A A A
t ,
i.e. Z ∞ 
uA
t = Et e−ρ(s−t) dcA
t+s .
t
Because the strategy is feasible, the process
  
A A A a − ι(qs )
dns = rns ds + (ks qs ) + g(qs ) + µs + σσs − r ds + ϕs (σ + σs )dZs − dcA
q q A q
s.
qs
stays nonnegative. Let ς = nB A A A A
t /nt , and consider the strategy ςks , ςdcs , ϕs ; s ≥ t of expert B.
This strategy is also feasible, because it leads to a non-negative wealth process nB A
t = ςnt , and it
delivers the expected utility of ςuA B A
t to expert B. Thus, ut ≥ ςut , leading to a contradiction.
Therefore, for all experts their expected utility under the optimal trading strategy is propor-
tional to wealth. It follows that θt = uA A B B
t /nt = ut /nt .

Proof of Lemma 2. First, assume that the process θt , t ≥ 0 represents marginal value of experts’
net worth. Let us show that then θt must satisfy the Bellman equation (5), which characterizes
the experts’ optimal strategies, and the transversality condition. Let {kt ≥ 0, dct ≥ 0, ϕt ≥ ϕ̃} be
an arbitrary admissible strategy (i.e. does not violate the solvency constraint). We argue that the
process Z t
Θt = e−ρs dcs + e−ρt θt nt
0
is always a supermartingale; and it is a martingale if the strategy {kt , dct , ϕt } is optimal. Note
that the maximal payoff that an expert can obtain at time t is
Z t+s 
−ρ(s0 −t) −ρs
θt nt ≥ Et e dcs0 + e θt+s nt+s ,
t

where equality is attained if the agent follows an optimal strategy from time t to t + s, since
θt+s nt+s is the maximal payoff that the agent can attain from time t + s onwards. Therefore,
Z t Z t+s 
−ρs0 −ρt −ρs0 −ρs
Θt = e dcs0 + e θt nt ≥ Et e dcs0 + e nt+s θt+s = Et [Θt+s ]
0 0

with equality if the agent follows the optimal strategy.

Differentiating Θt with respect to t using Ito’s lemma, we find
dΘt = e−ρt (dct − ρθt nt dt + d(θt nt ))
For the optimal strategy we have E[dct − ρθt nt dt + d(θt nt )] = 0 (since Θt is a martingale), and
for any arbitrary strategy we have E[dct − ρθt nt dt + d(θt nt )] ≤ 0 (since Θt is a supermartingale).
Therefore, the optimal strategy of any expert is characterized by the Bellman equation (5). To
verify that the transversality condition holds under an optimal strategy kt , ct , ϕt , note that (a) the
expected payoff of an expert with net worth nt is given by
Z ∞  Z t 
−ρs −ρs
θ0 n0 = E e dcs = lim E e dcs ,
0 t→∞ 0

41
where expectation value and limit can be interchanged by the Monotone Convergence Theorem
(because dcs ≥ 0), and (b) for all t,
Z t 
−ρs −ρt
θ0 n0 = E e dcs + e θt nt .
0

Taking t → ∞ in the latter formula, and combining with the former, we get the transversality
condition.
Conversely, let us show that if a process θt satisfies the Bellman equation and the transversality
condition holds, then θt represents the experts’ marginal value of net worth and characterizes their
optimal strategies. Note that, as we just demonstrated, equation (5) implies that the process Θt is
always a supermartingale, and a martingale for any strategy {kt , dct , ϕt } that attains the maximum
in equation (5). Thus, any expert who follows such a strategy attains the payoff of
Z ∞  Z t 
−ρs −ρs
E e dcs = lim E e dcs = lim (θ0 n0 − E[e−ρt θt nt ]) = θ0 n0
0 t→∞ 0 t→∞

where the last equality follows from the transversality condition.

Any alternative strategy achieves utility
Z t 
lim E −ρs
e dcs ≤ lim (θ0 n0 − E[e−ρt θt nt ]) ≤ θ0 n0
t→∞ 0 t→∞

where the last inequality holds because θt nt ≥ 0. We conclude that θ0 n0 is the maximal utility
that any expert with net worth n0 can attain, and that the optimal strategy must solve the
maximization problem in the Bellman equation.
Proof of Lemma 3. Aggregating over all experts, the law of motion of Nt is

dNt = rNt dt + ψt (Kt qt )[(Et [rtk ] − r) dt + ϕt (σ + σtq ) dZt ] − dCt ,

where Ct is are aggregate payouts, and the law of motion of Kt is

dKt = (ψt g(qt ) − (1 − ψt )δ)Kt dt + σKt dZt

d(1/Kt ) = −(ψt g(qt ) − (1 − ψt )δ)(1/Kt ) dt + σ 2 (1/Kt ) dt − σ(1/Kt ) dZt .

Combining the two equations, and using Ito’s lemma, we get

dηt = r − ψt g(qt ) + (1 − ψt )δ + σ 2 ηt dt + ψt qt (Et [rtk ] − r) dt − ψt ϕt qt σ(σ + σtq ) dt

+ (ψt qt ϕt (σ + σtq ) − σηt ) dZt − dζt

Substituting σtη = ψt ϕt qt (σ + σtq )/ηt − σ into the expression for the drift of ηt , we get (6).
Furthermore, if σtq ≥ 0, σtθ ≤ 0 and ψt > 0, then Proposition 1 implies that ϕt = ϕ̃, E[rtk ] − r =
−ϕ̃σtθ (σ + σtq ), and so

µηt = r − ψt g(qt ) + (1 − ψt )δ − σtη (σ + σtθ ) − σσtθ

Proof of Proposition 6. First, it is suboptimal to employ monitoring the event that the value
of the firm’s assets falls by less than nt− /ϕ̃ due to a jump. It is possible to guarantee that jumps

42
of size nt− /ϕ̃ or less are never caused by benefits extraction by subtracting value ϕ̃kt |dJti | from the
expert’s inside equity stake in the event that such a jump occurs. Such an incentive mechanism is
costless, since the expert is risk-neutral with respect to jump risks as they are uncorrelated with
the experts’ marginal value of net worth θt . At the same time, monitoring carries the deadweight
loss of a verification cost. Also, monitoring is not an effective way to prevent continuous diversion
of private benefits, because outside investors have to pay a positive cost of monitoring in response
to a possible infinitesimal deviation (recall that we disallow randomized monitoring).
Second, monitoring has to be employed in the event that the value of the firm’s assets falls by
more than nt− /ϕ̃, because it is the only way to prevent benefit extraction in such large quantities
(other than simply keeping the value of the assets below nt− /ϕ̃). Without loss of generality, we
can consider contracts that leave the expert with zero net worth if he is caught diverting such large
amounts for private benefit. In the event that a loss of size more than nt− /ϕ̃ is verified to have
occurred without benefit extraction, recovered capital can be split arbitrarily between the expert
and outside investors in an optimal contract. Because the expert is risk-neutral with respect to
idiosyncratic risks uncorrelated with aggregate shocks, without loss of generality we can assume
that all recovered capital goes to outside debt holders.24
In this case, to compensate outside investors for monitoring costs and for the expected value
lost in possible default (i.e. event when costly state verification is triggered), expert’s net worth
has to evolve according to
 
a − ι(qt ) q q
dnt = rnt dt + (kt qt ) + g(qt ) + µt + σσt − r − L(ϑt ) − C(ϑt ) dt
qt

i q
+dJt + ϕ̃(σ + σt ) dZt − dct ,

where ϑt = 1 − nt /(ϕ̃qt kt ) is the expert’s debt to total asset ratio (leverage). We set ϕt = ϕ̃ to
minimize the expert’s exposure to aggregate risk, because we assumed that σtq ≥ 0 and σtθ ≤ 0.

Appendix B: Contracting on kt
Appendix B analyzes the case in which contracting directly on kt is possible instead of kt qt . An
expert manages capital that follows

dkt = (Φ(ιt ) − δ − bt ) kt dt + σkt dZt .

where bt is the rate of private benefit extraction, and produces output (a − ιt )dt. Furthermore,
suppose that the expert can get the marginal benefit of ϕ̃ ≤ 1 units of capital per unit diverted.
Denote by qt the market price of capital, by θt , the value of expert funds per dollar, by ι(qt ) the
optimal level of investment and by g(qt ) = Φ (ι(qt )) − δ the implied growth rate. What is the
optimal contract, if kt rather than kt qt is used as the measure of performance? In this section we
follow the literature on dynamic contracting to derive the implications of contracting directly on
kt , e.g. see DeMarzo and Sannikov (2006).
24
There are other optimal contracts, for example the expert could be fully insured against drops in asset value
that are verified to involve no benefit extraction. Of course, the expert would have to pay a ‘premium’ for such
insurance in the event that there were no jump losses.

43
 Consider contracts based on the agent’s net worth as a state variable. The “official”net worth
follows
dnt = rnt dt + βt (dkt − g(qt )kt dt) − σtθ βt σkt dt, (8)
and the agent also gets funds at rate ϕ̃bt qt if he extracts benefits bt ≥ 0. The incentive constraint
is
βt ≥ ϕ̃qt
since the expert gets ϕ̃qt units of net worth (that can be used elsewhere to gain the utility of ϕ̃qt θt )
for one unit of capital diverted. Note that the stochastic as well as the deterministic portion of
the law of motion of nt depends directly on kt , so households need to observe kt directly in order
to write a contract that rewards the expert according to equation (8).
Note that e−ρt θt nt is a martingale when the expert refrains from extracting benefits and does
not consume. We have

d(θt nt ) = θt (rnt dt + βt σkt dZt − σtθ βt σkt dt) + (µθt dt + σtθ dZt )θt nt + σtθ θt βt σkt dt
= θt (rnt dt + βt σkt dZt ) + ((ρ − r) dt + σtθ dZt ) θt nt
= ρ(θt nt ) dt + volatility term,

where we use as in Section 3 the property that µθt = (ρ − r).

Next, we study the price of capital, qt . We derive a pricing equation for capital by setting the
expected return that households earn from investing in capital to r.
If contracting is based on kt only, then households hire experts to manage their capital, but
households themselves take on the price risk. The market price of capital still depends on the
experts’ risk-taking capacity. The return that households earn on their capital holdings kt is given
by

(kt qt ) rkt = (a − ι(qt )) kt dt + d (qt kt ) − βt kt σ dZt + βt σtθ σkt dt

= (a − ι (qt )) kt dt + (qt kt ) [(µqt + g (qt ) + σσtq ) dt + (σ + σtq ) dZt ]
−βt kt σ dZt + βt σtθ σkt dt

If σtθ < 0, then households optimally set βt = ϕ̃qt to minimize the costs of compensating experts
for risk. In expectation rkt should equal r, so we need

a − ι (qt )
+ µqt + g(qt ) + σσtq − r + ϕ̃σtθ σ = 0.
qt

This equation is different from the pricing equation (EK) because the risk premium is based only
on exogenous risk (for which households must compensate the experts that manage their capital).
Also, the law of motion of ηt will be different. Combining the law of motion of nt and the
condition that the households must get an expected return of r, we get the equation
  
a − ι(qt ) q q
dnt = rnt dt + (kt qt ) + µt + g(qt ) + σσt − r dt + ϕ̃σ dZt − dct ,
qt

which does not have the endogenous risk term. As a result,

  
a − ι(qt ) q q
dNt = rNt dt + ψt (Kt qt ) + µt + g(qt ) + σσt − r dt + ϕ̃σ dZt − dCt .
qt

44
Because dKt /Kt = (ψt g (qt ) − (1 − ψt ) δ) dt + σ dZt , and so

d (1/Kt ) / (1/Kt ) = − (ψt g(qt ) − (1 − ψt ) δ) dt + σ 2 dt − σ dZt ,

we get

r − ψt g (qt ) + (1 − ψt ) δ + σ 2 ηt dt


dηt =
 
a − ι(qt ) q q
+ψt qt + µt + g(qt ) + σσt − r dt
qt
−ψt ϕ̃qt σ 2 dt + (ψt ϕ̃qt − ηt ) σ dZt − dζt .

The volatilities of ηt and qt are found to be

q 0 (ηt ) η
 
η ψt ϕ̃qt
σt = −1 σ and σtq = σ ηt ,
ηt qt t
so there is still amplification through leverage, but no more feedback effect through prices.

Appendix C. Stationary Distribution

Suppose that Xt is a stochastic process that evolve on the state space [xL , xR ] according to the
equation
dXt = µx (Xt ) dt + σ x (Xt ) dZt (9)
If at time t = 0, Xt is distributed according to the density d(x, 0), then the density of Xt at all
future dates t ≥ 0 is described by the forward Kolmogorov equations:
∂ ∂ 1 ∂2 2
d (x, t) = − (µx (x) d (x, t)) + x


σ (x) d (x, t) .
∂t ∂x 2 ∂x2
If one of the endpoints is a reflecting barrier, then the boundary condition at that point is
1 ∂ x 2
−µx (x)d(x, t) + (σ (x) d(x, t)) = 0.
2 ∂x
A stationary density stays fixed over time under the law of motion of the process, so the
left-hand side of the Kolmogorov forward equation is ∂d(x,t) ∂t
= 0. If one of the endpoints of the
interval [xL , xR ] is reflecting, then integrating with respect to x and using the boundary condition
at the reflecting barrier to pin down the integration constant, we find that the stationary density
is characterized by the first-order ordinary differential equation
1 ∂ x 2
−µx (x)d(x) + (σ (x) d(x)) = 0.
2 ∂x
To compute the stationary density numerically, it is convenient to work with the function
D(x) = σ x (x)2 d(x), which satisfies the ODE

µx (x)
D0 (x) = 2 D(x). (10)
σ x (x)2
D(x)
Then d(x) can be found from D(x) using d(x) = σ x (x)2
.

45
 With absorbing boundaries, the process eventually ends up absorbed (and so the stationary
distribution is degenerate) unless the law of motion prevents (9) it from hitting the boundary with
probability one. A non-degenerate stationary density exists with an absorbing boundary at xL if
the boundary condition D(xL ) = 0 can be satisfied together with D(x0 ) > 0 for x0 > xL . For this
to happen, we need
Z x0 x 0
2µ (x ) 0
log D(x) = log D(x0 ) − dx → −∞, as x → xL
x σ x (x0 )2
R x x (x)
i.e xL0 2µ
σ x (x)2
dx = ∞. This condition is satisfied whenever the drift that pushes Xt away from the
boundary xL (so we need µx (x) > 0) is strong enough working against the volatility that may move
Xt towards xL . For example, if Xt behaves as a geometricR Brownian motion
R x0 near the boundary
x x x0 2µx (x) 2µ
xL = 0, i.e. µ (x) = µx and σ (x) = σx, with µ > 0, then 0 σx (x)2 dx = 0 σ2 x dx = ∞.

46