Gordo Ps

Download as ps, pdf, or txt
Download as ps, pdf, or txt
You are on page 1of 74

DYNAMIC HEDGING WITH TRANSACTION COSTS:

FROM LATTICE MODELS TO NONLINEAR VOLATILITY


AND FREE-BOUNDARY PROBLEMS

Antonio Paras (1 ; 2) and Marco Avellaneda (1 ; 3)

Abstract. We study the dynamic hedging of portfolios of options and other deriva-
tive securities in the presence of transaction costs. Following Bensaid, Lesne, Pages &
Scheinkman (1992), we examine hedging strategies which are risk-averse and have the
least initial cost, in the framework of a multiperiod binomial model. This paper considers
the asymptotic limit of the model as the number of trading periods becomes large.p This
limit is characterized in terms of nonlinear di usion equations. If A = k=( dt) < 1
(k is the roundrip transaction cost,  is the volatility and dt is the lag between trading
dates), the optimal cost approaches the solution of a nonlinear Black-Scholes-type
p equa-
p in which the volatility is dynamically adjusted upward to  1 + A or downward to
tion
 1 A according to the local convexity of the solution. For A = 1, the upward ad-
justment is similar but the downward adjustment assigns zero nominal volatility to the
underlying asset for long-Gamma positions. In the latter case, the optimal cost function
is the solution of a free-boundary problem. We also characterize the associated hedging
strategies. We shown that if A < 1, it is optimal to replicate the nal payo via \nonlinear
Delta hedging". On the other hand, if A = 1, the optimal strategies are path-dependent,
non-unique and typically super-replicate the nal payo .

(1)
Courant Institute of Mathematical Sciences, New York University, N.Y.,N.Y.,10012.
(2)
These results were obtained as part of the rst author's doctoral dissertation.
(3)
Partially supported by the Institute for Advanced Study, Princeton, N.J.,08540 .
1
1. Introduction and Statement of the Main Results

The problem of accounting for transaction costs in the dynamic hedging of derivative
securities has attracted considerable attention from theoreticians and practitioners. S-
trategies which account for transaction costs and pricing models that incorporate the costs
into the premium have considerable practical interest, especially for trading derivatives in
markets with moderate or low liquidity.
The problem can be formulated in terms of an agent that buys and sells options on the
stock of a company. At some point in time, he or she decides to hedge the \book", or
options portfolio, against future price uctuations. The agent would like to determine the
least costly strategy, taking into account the projected transaction costs due to dynamic
hedging. The initial cost of such strategy can be interpreted as the minimal capital reserve
needed to protect the portfolio against future market moves.1
We shall make the assumption that the agent is totally risk-averse: only strategies
which ensure non-negative cash- ows after closing all positions are deemed admissible.
This assumption is important in the framework of this paper but by no means necessary.
Several theorists have considered hedging strategies which allow for losses but maximize a
utility function assigned to the agent (Constantinides (1979,1986), Hodges and Neuberger
(1989), Davis, Panas and Zariphopolou (1993)). In principle, a utility-based approach
may present greater exibility2 but has the disadvantage of producing utility-dependent
results. For this and other reasons, total risk-aversion plays a central role in the assessment
of transaction costs in derivative strategies.
Risk-averse hedging in the presence of transaction costs was rst considered by Leland
(1985) for the log-normal model and later by Boyle and Vorst (1992) for the binomial
lattice. Both papers consider only the problem of hedging a single option. A non-trivial
generalization of the Leland model applicable to option portfolios and exotic options was
proposed later by Hoggard, Whalley and Wilmott (1993), based on dynamic replication of
the payo . However, Boyle and Vorst and Hoggard et al. both observed that replication
may not always be feasible since it can lead to in nite option prices when transaction costs
exceed a critical value and the agent is long Gamma (as in the case of a long call position).
Avellaneda and Paras (1994) examined this issue, which is related to the mathemat-
ical ill-posedness of the replication equation. They proposed an explanation for the ill-
posedness, which points to a fundamental di erence between long and short options posi-
tions in markets with large bid-o er spreads. The risk-averse seller of an option is obliged
to dynamically hedge his or her exposure in the cash market, regardless of transaction
1 Throughout this paper, we consider only dynamic hedging with shares of stock and a money-market
account. Of course, in practice, traders also hedge their books with other derivative securities to (among
other things) diminish the transaction costs. We assume that the agent has already taken a de nite
position in the derivatives market and seeks to hedge the \residual" exposure with a position in the cash
market.
2
The utility framework contains risk-aversion as a special case in which losses are assigned \in nitely
negative utility".
2
costs. On the other hand, the buyer of the option risks only the initial premium: hedging
is done primarily to o set the time-decay in the option's value. It is intuitively clear (and
can be proved mathematically) that if transaction costs in the cash market are suciently
large, Delta-hedging to o set time-decay is impossible. This key observation applies also
to option portfolios: if transaction costs are high and volatility is low, Delta-hedging a
position which is long-Gamma is counterproductive. In such situations, a temporary hold-
ing strategy is preferable, since it does not carry immediate market risk as long as capital
reserves are suciently high. Based on this observation, Avellaneda and Paras proposed a
new scheme for pricing and hedging option portfolios which is based on solving an obsta-
cle problem for Leland's volatility-adjusted PDE. The corresponding dynamic hedges are
non-Markovian (path-dependent) and dominate { or super-replicate { the nal payo .
The aforementioned strategies are based on replication or super-replication. This raises
the question of characterizing strategies which minimize the initial cost of hedging. How
are the two concepts related ? The least-initial-cost formulation was rst considered by
Bensaid, Lesne, Pages and Scheinkman (1992) (BLPS) in the framework of a nite-period
binomial tree, using a a dynamic programming algorithm (see Section 2, equation 2.10).
Due to dependency on the previous stock holdings, the BLPS optimal hedging strategies
are path-dependent, with some notable exceptions. For instance, in the case of a short
option settled in shares, Bensaid et al. showed that it is always optimal to replicate the
payo . This result has an interesting consequence: the BLPS least-initial-cost strategy
for a short option corresponds, in the asymptotic limit of many trading periods, to a
Black-Scholes pricing formula with volatility adjusted upwards to re ect transaction costs,
analogous to the Boyle and Vorst (1992) formula.3
The least-initial-cost pricing and hedging of complex options portfolios is the main
subject of this paper. Our approach consists in characterizing the solutions of the BLPS
algorithm in the limit of in nitely many trading periods for general contingent claims. We
show that the algorithm admits a simple interpretation in this asymptotic limit. In fact,
the optimal initial costs satisfy nonlinear partial di erential equations analogous to those
proposed by Hoggard et al. and by Avellaneda and Paras, with minor changes in the values
of the adjusted volatilities that re ect the di erence between the normal and binomial
statistics. As a consequence of this result, we determine precisely in which instances the
BLPS algorithm gives rise to replicating strategies and when the optimal strategies are
path-dependent and dominating. We also show that whenever path-dependency holds for
BLPS, the hedging strategies are essentially analogous to those proposed by Avellaneda and
Paras (1994) for the log-normal model. Thus, path-dependency occurs when (i) transaction
costs exceed a critical level and (ii) the agent is long-Gamma for some level of the spot
price.

1.1 Overview of the results


Following Bensaid, Lesne, Pages and Scheinkman (1992), we formulate the problem in
3
This is also the analogue of the Leland (1985) for the binomial tree.

3
terms of the binomial probability model. In this framework, there are N +1 trading dates,
the last one being the expiration date. The lag between successive trading dates is xed.
The impact of imperfect liquidity is modeled through a bid-o er spread in the cash market,
assuming that the agent will buy stock at the o er and sell at the bid. The bid/o er prices
for one share of stock are de ned as
   
S bid = Sn k
1 2 and S ask = Sn k
1+ 2 ;
n n

where Sn is the average between the bid and o er and k represents the (percentage) round-
trip transaction cost for buying/selling stock, i.e.,4

k = Sn S Sn :
ask bid

For simplicity, we assume that k is constant. The change in the average stock price over a
single period is modeled by a two-state random variable

p Sn U
Sn
1-p Sn D

with probabilities p and 1 p for upward and downward moves. We also assume that the
dollar return for lending/borrowing over one trading period is R (constant), and that U ,
D and R satisfy the \pure" no-arbitrage condition
D < 1+R < U :

A key feature of this model is given in


De nition 1.1. The stock is said to be risk-feasible if there is a positive probability of
posting a pro t by following either one of the following strategies:
(i) Borrow money to purchase a share of stock; hold the stock for one trading period;
sell the stock and pay back the loan.
(ii) Short-sell a share of stock; deposit the proceeds in a money market account for one
period; close the account and unwind the short position after the period.
4
Thus, an agent who buys and immediately sells one share of stock assumes a loss of k  Sn dollars.

4
Mathematically, conditions (i) and (ii) are satis ed if and only if
   
k=2
D 11 + k= < 1 + R < U 11 + k=2 :
2 k=2

The risk-feasibility of an asset depends on how much its price is expected to oscillate
over a single trading period in relation to the round-trip transaction costs. As we shall see,
the BLPS optimal strategies are very di erent according to whether risk-feasibility holds
or not. If the stock is risk-feasible, replicating strategies are always optimal for arbitrary
contingent claims. Otherwise, replication may not be optimal.
To illustrate the sub-optimality of replicating strategies, we consider a simple example.
Example 1.2 Assume there is a single trading period. An agent wishes to hedge a
short position in a hypothetical claim contingent on the value of a non-risk-feasible stock.
Assume that this claim pays $1 in the \up" state and $0 in the \down" state. Consider
rst the case in which the agent has no endowment in shares. Opening and subsequently
closing any position in shares after one period will result in an overall loss under both states
of the world. Hence, it follows that the initial hedging cost of any replicating portfolio
using shares and an interest-bearing money-market account must exceed the present value
of $1 (the maximum liability). On the other hand, simply holding the present value of
$1 (i.e., $(1 + R) 1 ) in cash constitutes a (cheaper) dominating riskless strategy. Notice,
however, that the situation is di erent if either (i) the agent has an endowment in shares
or (ii) if the contingent claim is settled in shares rather than in cash. In such cases, a
replicating strategy may be cheaper than $(1 + R) 1 because transaction costs are lower
for the agent. As a matter of fact, the reader can verify that it is optimal to replicate when
the agent's endowment in shares exceeds 1=[S0(1 k=2)(U D)] shares in Case (i) and
1=[S0(U D)] in Case (ii) (See Section 3). This suggests that, aside from the risk-feasibility
of the underlying cash instrument, the agent's endowment and the form of payment may
determine whether the it is optimal to replicate or not.
For a binomial model with a large number N of trading periods per year | and hence
with small duration between successive trading dates dt = T=N  1 | risk-feasibility is
equivalent to

A  pk < 1; (1.1)
 dt
where  is the annualized stock volatility. We shall refer to A as the Leland number.
Table 1.1 shows that the Leland number takes moderate values (between 0 and 10) for
standard values of volatility, time-lag and transaction costs ( volatility of 20%, time-lags
ranging from one week to four times a day and round-trip costs of up to 6%). From this
table, we draw the heuristic conclusion that if the time-to-expiration is suciently long,
then (i) the minimum lag between trades can be taken to be small enough to warrant a
5
k/2 0.01 0.02 0.03
dt
one week
-2 0.72 1.44 2.17
1.9*10
half week
-3 1.02 2.04 3.06
9.6*10
one day
-3 1.61 3.23 4.84
3.8*10
half day
-3 2.28 4.57 6.85
1.9*10
quarter day
-4 3.23 6.46 9.69
9.6*10
Table 1.1. Leland number A for di erent values of the minimum lag between trading
dates dt (from one week to 1=4 day) with round-trip transaction costs between 2% and 6%
and volatility of 20%.
continuum description of the system, and (ii) at the same time, A can be considered to be
nite. The interesting asymptotic regime of parameters for the asymptotic analysis of the
BLPS algorithm is, in fact, dt  1 and A = O(1).5
Under these assumptions, the asymptotic analysis of the BLPS algorithm in the limit
N % 1 yields the following results:
(i) If the underlying asset is risk-feasible, the minimal initial cost of risk-averse hedg-
ing a European-style contingent claim with payo value F (S ) and expiration date
T approaches uniformly to the solution of the nonlinear di usion equation
@P + 1 S 2 b2  @ 2 P   @ 2P + r S @P P

= 0; (1.2)
@t 2 @S 2 @S 2 @S
with
P (S; T ) = F (S ) ;
5
The regime dt  1, A  1, which might be considered relevant when dt is small, gives rise to
trivial buy/hold strategies. The BLPS cost function for this regime is recovered by rst letting dt % 1
and then taking the limit A  1 in the asymptotic equations (1.4) derived hereafter (cf. Avellaneda and
Paras (1994)).
6
where  2    2 
b2 @@SP2 =  1 + A sign @@SP2 :
2 (1.3)

Here, r represents the annualized interest rate. Notice that, since the Leland number is
less than unity, the nonlinear parabolic equation (1.2) is well-posed. The optimal hedging
strategy corresponds to maintaining a hedge-ratio of @P @S shares of stock at each trading
date.
(ii) If the underlying asset is not risk-feasible, the minimum initial cost for hedging a
European-style contingent claim with payo value F (S ) expiring at time T can be
approximated as N % 1 by the solution of the \obstacle problem":

P (S; t)  e r(T t) F (er(T t) S ) ;

@P + 1 S 2 2 (1 + A)  @ 2P + r S @P
 

@t 2 @S 2 @S P 5 0

for all (S; t) and

@P + 1 S 2 2 (1 + A)  @ 2 P + r S @P
 

@t 2 @S 2 @S P = 0

if P (S; t) > e r(T t) F (er(T t) S ),


with nal condition
P (S; T ) = F (S ) : (1.4)

This obstacle problem reduces to a linear PDE analogous to Leland's equation if FSS is
positive for all S . Otherwise, the problem splits the (S; t)- plane into two regions: one
where the PDE is satis ed, and another, \the contact set", in which the value function
coincides with the \obstacle" e r(T t) F (er(T t) S ) and the nominal volatility is zero. The
corresponding dynamic hedging strategies alternate between delta-hedging and holding
periods without hedge adjustments. The times at which the agent switches from delta-
hedging to static hedging are determined by the position of the spot price with respect to
the contact set and by the Delta of the agent's portfolio.
In simple words, the asymptotic analysis of the BLPS algorithm shows that the impact
on the total hedging cost arising from
p bid-o er spreads can be taken into account by
adjusting the volatility upward to  1 + A when the agent ispshort \nonlinear Gamma"
( @ 2P=@S 2 > 0 ) and adjusting the volatility downward to  1 A or to zero when the
7
Short Call Position.
V($) V($)
A = 1.12 A = 0.94
0.2 0.2

0.15 0.15

0.1 0.1

0.05 0.05

0 0
0.8 0.9 1 1.1 S($) 0.8 0.9 1 1.1 S($)

V($) A = 0.71 V($) A = 0.50


0.2 0.2

0.15 0.15

0.1 0.1

0.05 0.05

0 0
0.8 0.9 1 1.1 S($) 0.8 0.9 1 1.1 S($)

Figure 1.1. Value of a short call position with strike price K = 1 with 6 months to expira-
tion. The parameter values are k = 1%,  = 20% and dt = 0:01 (A = 0:5); 0:005 (A =
0:71); 0:003(A = 0:94) and 0:02 (A = 1:12). The curves corresponding to the PDEs
and the BLPS algorithm (thick line) are visually indistinguishable. The dotted lines represent
the intrinsic value of the call.
agent is long \nonlinear Gamma" (@ 2 P=@S 2 5 0).6 When A = 1, an adjustment to zero
volatility should be made in the latter case.
In addition to providing a characterization of least-initial-cost dynamic hedging, the
asymptotic PDEs provide an ecient computational procedure for solving approximate-
ly the BLPS algorithm. In fact, BLPS involves two state-variables | price and stock
Since P represents costs the agent's cost (liabilities), the pro t-loss is P . Hence our de nition of
6

Gamma.
8
Long Call Position.
V($) A = 1.12 V($) A = 0.94
0 0

0.05 0.05

0.1 0.1

0.15 0.15

0.2 0.2
0.8 0.9 1 1.1 S($) 0.8 0.9 1 1.1 S($)

V($) A = 0.71 V($) A = 0.50


0 0

0.05 0.05

0.1 0.1

0.15 0.15

0.2 0.2
0.8 0.9 1 1.1 S($) 0.8 0.9 1 1.1 S($)

Figure 1.2. Same as Figure 1.1 for a long call position. Again, the agreement between
BLPS and the nonlinear PDE s is remarkable. Notice that the long call position is priced at
intrinsic value (as if the volatility were zero) for A = 1:12.

holdings | (see equation (2.11) in Section 2), while the PDEs involve only the spot price.
More importantly, the BLPS scheme has exponential complexity in N when the underlying
asset is not risk-feasible, unlike the nite-di erence PDE solvers which are O(N 2 ). The
error which arises from the use of a nite-di erence scheme for thep PDEs (1.2) or (1.4)
to approximate the BLPS algorithm is essentially proportional to dt, and hence small.
The quality of the approximation is exhibited in Figures 1.1, 1.2, 1.3 and 1.4 for various
contingent claims under realistic values of the model parameters for equity markets.
To conclude this Introduction, we address an important issue regarding risk-feasibility
9
Short Butter y Position.
V($) A = 1.12 V($) A = 0.94
0.1 0.1

0.08 0.08

0.06 0.06

0.04 0.04

0.02 0.02

0 0
0.8 0.9 1 1.1 0.8 0.9 1 1.1
S($) S($)

V($) A = 0.71 V($) A = 0.50


0.1 0.1

0.08 0.08

0.06 0.06

0.04 0.04

0.02 0.02

0 0
0.8 0.9 1 1.1 0.8 0.9 1 1.1
S($) S($)

Figure 1.3. Comparison between the value functions obtained from the di erential equa-
tions and the Bensaid-Lesne-Pages-Scheinkman algorithm for a short 0:9 1:0 1:1 butter y
spread with six months to expiration. The smooth curves are the solutions of the PDEs (1.2)
or (1.4) and the oscillating curves are solutions of the BLPS algorithm. The dotted lines
represents the intrinsic value of the spread in present dollars. The parameter values are as in
Figure 1.1.

and risk-aversion. As shown above, the initial cost and the strategy to follow are deter-
mined by the Leland number A or equivalently, by the minimum lag between adjustments,
dt. The question of how to choose the Leland number in practice poses itself naturally.
Clearly, the BLPS model does not address this issue, requiring an endogenous speci cation
of dt. In practice, a choice of the mesh-size cannot be made without taking into account
the agent's aversion to hedge slippage between trading dates. This remark also applies to
10
Short Digital Position.
V($) A = 1.12 V($) A = 0.94
0.1 0.1

0.08 0.08

0.06 0.06

0.04 0.04

0.02 0.02

0 0
0.85 0.9 0.95 0.98 1 1.05 0.85 0.9 0.95 0.98 1 1.05
S($) S($)

V($) A = 0.71 V($) A = 0.50


0.1 0.1

0.08 0.08

0.06 0.06

0.04 0.04

0.02 0.02

0 0
0.85 0.9 0.95 0.98 1 1.05 0.85 0.9 0.95 0.98 1 1.05
S($) S($)

Figure 1.4. Same as Figure 1.3 for a digital option with 6 months to expiration. The
payo is $1 if the stock is worth more than $1 at expiry and $0 otherwise.

the asymptotic PDEs. This dependency on the time-lag is a conceptual disadvantage of the
model vis a vis the utility-dependent approaches such as Davis, Panas and Zariphopolou
(1993) which determine dt endogenously. It is clear however that small Leland numbers
correspond to large intervals between adjustments, with greater slippage risk, and large
Leland numbers correspond to smaller lags between adjustments and thus to less risk but
more transaction costs. A practical way to specify dt would be to view the slippage risk as
being determined by the magnitudes of the nonlinear Delta (@P=@S ) and nonlinear Gam-
ma ( @ 2 P=@S 2). Thus, a choice of dt can be made using either a \worst-case
2P
scenario"
time-step, or a time-step that varies according to the magnitudes of @S , @S2 and the time
@P @
to expiration. In practice, speci cation of dt can be done by analyzing pro t/loss his-
11
tograms obtained by Monte Carlo simulation (Avellaneda and Paras (1994)). The overall
result would be to have a variable Leland number A and a PDE that would combine the
features of (1.2)-(1.3) and (1.4).7

The remaining sections of this of the paper contain a detailed mathematical analysis
of the passage from the binomial lattice model to the asymptotic PDEs. In Section 2, we
review the BLPS algorithm. In Section 3, we study two special conditions under which the
optimal hedging strategy is to replicate the payo and thus BLPS reduces to a backward-
induction algorithm on the binomial lattice. This analysis pertains to the properties of
the BLPS algorithm at the discrete level. In Section 4, we study the asymptotic behav-
ior of the algorithm in the cases when replicating strategies are optimal. The technique
used in the proofs is to derive approximate solutions of the backward-induction algorithm
using the solutions of the PDE (1.2). This requires some regularity properties of the so-
lutions as well as a \comparison principle", adapted to the nonlinear induction relation,
to evaluate rigorously the approximation error. Section 5 studies the case A = 1, when
optimal dynamic hedging is path-dependent. To characterize this regime, we construct
super-replicating strategies using the solution of the obstacle problem following Avellane-
da and Paras (1994). This leads to an upper bound on the BLPS solution in terms of
(1.4). A lower bound in terms of P (S; t) is obtained by comparing the BLPS optimal cost
with the values of certain \barrier options" that knock out on the boundary of the contact
set of the obstacle problem. The main asymptotic results are contained in Theorems 4.4,
4.5 and 5.4 in Sections 4 and 5. For the reader's convenience, the regularity properties of
equations (1.2) and (1.4) used in the proofs are presented in an Appendix.

7The use of variable-dt strategies was suggested in Whalley and Wilmott (1993) for small values of
A for which equation (1.3) remains well-posed. More recently, Grannan and Swindle (1995) investigated
strategies for hedging standard options using Leland's framework with variable time-steps. Other recent
papers which investigate option replication with transaction costs and the time-lag issue include Hodges
and Clelow (1993,1994), Flesaker and Houghston (1994), Albanese and Tompaidis (1995) and Cvitanic
and Karatzas (1995).
12
2. The Bensaid-Lesne-Pages-Scheinkman Algorithm

Consider a binomial tree with N periods, starting at the initial date t0 and ending at
the expiration date tN . A trading strategy is de ned as a sequence of portfolios (n; Bn),
n = 0; . . . ; N , to be held during each period. Here, n represents the number of shares of
the underlying security held long or short and Bn the balance of a money-market account.
Adjustments of the portfolio are done at the beginning of each trading period: at time tn
an amount n n 1 of shares is traded and the balance of the operation (transaction cost
included) is added to the money-market account. A self- nanced strategy is one for which
the balance in the money-market account after adjusting the share holdings is exactly
equal to Bn . We will only consider self- nanced strategies.
At time tn, but before readjusting the position, the portfolio is composed of n 1 shares
and Bn 1(1 + R) in cash. The cost of trading n n 1 shares is

(n n 1)Sn + k2 jn n 1jSn  (n n 1)Sn;

where 8
>
< (1 + k2 ) y ; y = 0
(y) = >
:
(1 k2 ) y ; y < 0:
Therefore, a self- nanced strategy satis es
Bn = Bn 1(1 + R) (n n 1)Sn ; n = 1; 2; :::; N: (2.1)
Applying this relation recursively, we obtain
n
X
Bn = B0(1 + R) n (j j 1)Sj (1 + R)n j ; (2.2)
j =1

so that at time tN
N
X
BN = B0(1 + R) N (j j 1 )Sj (1 + R)N j : (2.3)
j =1

This formula shows that a self- nanced strategy is uniquely determined by the initial
money-market balance B0 and the (projected) sequence of stock holdings f0; . . . ; N g.8
A European-style derivative security is a contingent claim with payo
8
It is assumed throughout the paper that the sequence fn g de ning a trading strategy is non-
anticipating with respect to the ow of information, i.e., n = n (S0 ; S1 ; :::;Sn ).
13
N = N (SN ); BN = BN (SN )
in shares and cash, respectively, at time tN . A dominating (or risk-averse) strategy of a
short position in this claim is such that
N = N and BN = BN (2.4)
for all nal states. Note that if a dominating strategy satis es N > N for some nal
state, the excess of shares can be sold and the pro t credited to the money-market account.
Therefore, we can restrict our attention to risk-averse hedging strategies for which
N = N ; BN = BN : (2.5)
If a hedging strategy satis es the more stringent conditions
N = N ; BN = BN (2.6)
for all nal states, we say that it is a replicating strategy.
Given an adapted sequence of stock holdings f0; . . . ; N = N g, it follows from (2.3)
that
8 9
< N
X =
B0 = maxN B  (1 + R) N + (j j 1 )Sj (1 + R) j (2.7)
N
fSj gj=1 :
j =1
;

is the minimum initial cash reserve required to constitute a dominating strategy of the
contingent claim (N ; BN ).9
Assume that at time t0 an agent is short the contingent claim and has an inven-
tory of  1 shares (long or short). The minimum investment required to implement
f0; . . . ; N = N g as a risk-averse hedging strategy is given by
 1S0 + (0  1)S0 + B0 ;
which represents the combined sum of the value of the initial endowment in shares, the
cost of trading 0  1 shares, and the minimum initial cash reserves necessary to
ensure that the strategy fj gNj=0 dominates the nal payo . An optimal hedging strategy
is one that, given the initial endowment  1, minimizes the sum of the last two terms,
(0  1)S0 + B0 . By de nition, this is the e ective cost of the strategy.
9
Here, the maximum is taken over all forward paths f Sj gNj=0 followed by the stock price.

14
According to this de nition, the minimum e ective cost over all possible risk-averse
hedging strategies (minimum e ective cost, for short) is

V0 ( 1; S0) = minN f (0  1)S0 + B0g


fj g j=0
8 9
< N
X = (2.8)
= minN max
N B (1 + R)
: N
N + (j j 1)Sj (1 + R) j :
fj gj=0 fSj gj=1 j =0
;

2.1 Dynamic Programming Equation.


A similar argument shows that if an agent who has an endowment of n 1 shares at time
tn sells the derivative security with payo (N ; BN ) and plans to implement the strategy
fj gNj=n , the minimum cash reserve she requires to generate a dominating strategy is

8 9
< N
X =
Bbn = max
N BN (1 + R) (N n) + (j j 1 )Sj (1 + R) (j n); : (2.9)
fSj gj=n+1 : j =n+1

Hence, the minimum e ective cost at time tn is

Vn(n 1; Sn) = minN fBbn + (n n 1)Sng


fj g j=n
8 9
< N
X =
= minN max B  (1 + R) (N n) + (j j 1)Sj (1 + R) (j n) ; ; (2.10)
N
fj gj=n fSj gj=n+1 : N
j =n

for n = 0; . . . ; N .
Proposition 2.1 The function Vn (; S ) satis es the equation

Vn(n 1; Sn) =
 

 (n
min n 1)Sn + 1 +1 R maxfVn+1(n; SnU ); Vn+1 (n; SnD)g (2.11)
n
15
with nal condition 10

VN (N 1; SN ) = BN (SN ) + (N (SN ) N 1)SN : (2.12)

Proof: The minimum cash reserve in eq. (2.9) can be rewritten in the form
( )
Bbn = S =max Bbn+1 + 1 (
n+1 n)Sn+1
n+1 fSn U;Sn Dg 1 + R 1 + R

1
= 1 + R fmax
n o
Bn+1 + (n+1 n)Sn+1 :
b
U;Dg

Therefore, we have
1 maxfBb + (
min Bbn = min n+1 n )Sn+1g
N
fj gj=n+1 fj gj=n+1 1 + R U;D n+1
N

1 max min fBb + (


= n+1 n)Sn+1 g
1 + R U;D fj gNj=n+1 n+1

= 1 +1 R maxfVn+1(n; SnU ); Vn+1 (n; SnD)g:

Hence, substituting this into (2.10), we conclude that

Vn(n 1; Sn) = minN f (n n 1)Sn + Bbng


fj gj=n
 
1
= min (n n 1 )Sn +
n 1 + R maxfVn+1(n; SnD); Vn+1 (n; SnD)g : (2.13)

We claim that the opposite inequality also holds. In fact, recall that Vn+1(n; Sn+1)
represents the minimum e ective cost at time tn+1 given that the endowment is n and
the spot price is Sn+1. Because of this, the maximum of the two amounts
10
The algorithms for solving the BLPS equation can have complexity O(e N ) with > 0 for some
contingent claims. This is due to increasing complexity with the time-to-maturity of the quantity
maxfVn+1 (n ;Sn U ); Vn+1 (n ; Sn D)g as a function of n . This function must be stored in memory
to deduce Vn from Vn+1 .
16
(n n 1)Sn + 1 +1 R Vn+1(n; SnU )

and

(n n 1)Sn + 1 +1 R Vn+1(n; SnD)

is sucient to cover the value of a riskless hedge at time tn regardless of whether


Sn+1 = SnU or SnD. Consequently, we have

Vn(n 1; Sn) 5
 
min
 (n n 1)Sn + 1 +1 R maxfVn+1(n; SnU ); Vn+1(n; SnD)g : (2.14)
n

Combining (2.13) and (2.14), we obtain the nonlinear dynamic programming equation
(2.10). The nal condition in (2.11) follows from the de nition of e ective cost. Q.E.D.
Remark 2.2. The argument generalizes to derivative securites which deliver11a sequence
of payo s at di erent dates, contingent on the value of the stock at each date. Any such
derivative security can be speci ed by its sequence of payo s (n; Bn ), where

n = n(Sn) ; Bn = Bn (Sn): (2.15)


The e ective cost function and the optimal hedging strategy for this security can be found
by solving the modi ed dynamic programming equation
Vn(n 1; Sn) =
 
min Bn + (n + n
n n 1)Sn + 1 +1 R maxfVn+1(n; SnU ); Vn+1 (n; SnD)g

(2.16)

with nal condition (2.12). Also, the same equation applies to derivative securities in
which the nal date tN is a random stopping time, like barrier options, since the proof of
Proposition 2.1 does not use the fact that N is deterministic.
11
Examples of such \contingent claims" include portfolios of European options with di erent maturities
and equity-linked debt instruments with coupon payments contingent on the price of a stock or a stock
index.
17
Remark 2.3 In the formalismpresented here, forward contracts and bonds correspond to
contingent claims such that N and BN are constant. It is easy to verify that for these
claims with no optionality the unique optimal strategy is then n = N for all n. The
minimum e ective cost at time tn is given by

Vn(n 1; Sn) = (N n 1) Sn + (1 + R1)N n BN


 
= N Sn + 1
(1 + R)N n B
N + k2 jN n 1j Sn n 1 Sn :
(2.17)

Here, the term in brackets corresponds to the well-known cost-of-carry formula (without
transaction costs). The other two terms represent the transaction cost of the initial share
purchase/sale minus the value of the agent's endowment.
More generally, assume that the nal payo has the form N = N (SN ) + (0) and
BN = BN (SN ) + B(0) where (0) and B(0) are constants. It is then easy to verify that
if [fngNn=0; B0 ] is an optimal strategy for hedging the nal payo (N (SN ); BN (SN ))
given initial endowment  1, then [fn + (0)gNn=0; B0 + B(0) (1 + R) N ] is an optimal
strategy for hedging the payo (N ; BN ) given an initial endowment  1 + (0).

2.2 A stability property of the BLPS algorithm.


Replication of the nal payo can lead to mathematical instabilities if transaction
costs are large (Whalley and Wilmott (1993), Boyle and Vorst (1992), Avellaneda and
Paras (1994) ). Even though there are nitely many trading periods, replication may be
more expensive than an (already expensive) buy-and-hold strategy. The BLPS algorithm
does not have this pathology.
Proposition
  2.4: Suppose that two European-style derivative securities with payo s
(1) (1) and (2); B (2) are such that their respective e ective costs at expiration
N ; BN N N
satisfy

VN(1) (; SN ) 5 VN(2) (; SN ) (2.18)


for all  and all SN . Then
Vn(1) (; Sn) 5 Vn(2) (; Sn) (2.19)
for all 0 5 n 5 N ;  and Sn.
18
Proof: Using the inequality (2.18) and equation (2.11), we nd that for n = N 1, we
have

VN(1) 1 (; SN 1) 5 VN(2) 1 (; SN 1) :


This argument can be applied recursively to obtain (2.19) for all n. Q.E.D.

19
3. Replication vs. Super-replication:
\Fine Structure " of the BLPS Algorithm

In this section, we show that the risk-feasibility of the underlying asset is sucient to
guarantee that replicating strategies are optimal. The dynamic programming equation
(2.11) (or (2.16)) then reduces to a backward-induction relation for the pairs (n; Bn)
(see Boyle and Vorst (1992)). Another condition under which replication is optimal is
the k-convexity of the nal payo . This assumption can be seen as a discrete version of
convexity of the value the nal payo .
Prices and state-variables corresponding to di erent nodes of the binomial tree are
represented using double-index notation; e.g. Snj represents the price at time tn at the
node (n; j ).
3.1 Risk-Feasibility.
Proposition 3.1. If the underlying asset is risk-feasible, i.e., if
1 + k=2 D < 1 + R < 1 k=2 U ; (3.1)
1 k=2 1 + k=2
then
(i) the optimal sequences of stock holdings fng and of cash balances fBng are
independent of the initial endowment  1;
(ii) the optimal hedging strategy is replicating;
(iii) given the pairs (jn+1; Bnj +1) and (jn+1 j +1
+1; Bn+1 ), the hedge ratio jn is the unique
solution of the equation

(jn+1 j j +1 j +1
+1 n)Sn+1 + Bn+1 = (n+1 n )Sn+1 + Bn+1 ;
j j j j
(3.2)

and the money-market account balance Bnj is given by

Bnj = 1 +1 R ( (jn+1 j +1 j +1
+1 n )Sn+1 + Bn+1 )
j

= 1 +1 R ( (jn+1 jn)Snj +1 + Bnj +1) : (3.3)

20
Proof. Assume rst that the starting date is tN 1, that the starting state is j ; ( i:e: S =
SNj 1) and that the initial endowment is N 2. We abbreviate the notation, setting

SN 1 = SNj 1 ;

SNU = SN 1U = SNj+1 ;

SND = SN 1D = SNj ;

(UN ; BNU ) = (jN+1; BNj+1 ) ;

and

(DN ; BND ) = (jN ; BNj ) :

According to Equation (2.11), the optimal hedge ratio N 1 is found by solving the
minimization problem

min max fU (N 1); D (N 1)g (3.4)


N 1

where the functions U and D are de ned by

U ()  ( N 2)SN 1 + 1 f (UN )SN 1U + BNU g (3.5)


1+R

and

D () = ( N 2)SN 1 + 1 +1 R f (DN )SN 1D + BND g: (3.6)

We claim that U () and D () are, respectively, strictly decreasing and strictly in-
creasing functions of . To see this, we di erentiate formally Equations (3.5) and (3.6).
Accordingly,
21
   
dU () = 1  k S 1 1 k S U
N 1
d 2 1+R 2 N 1
   
5 1+
k S 1 1 k S U
N 1
2 1+R 2 N 1
   
k (1 k= 2) U
= 1 + 2 SN 1 1 (1 + k=2)(1 + R)

<0;

where risk-feasibility was used to obtain the last inequality. A similar calculation shows
that

dD () = 1  k  S  
1 1 k S D
d 2 N 1 1+R 2 N 1
   
k 1 k
= 1
2 SN 1 1 + R 1 + 2 SN 1D
   
= 1 k2 SN 1 1 (1 +(1R+)(1
k=2)D
k=2)
>0:

Since U is decreasing and D increasing, we have


U () for  < N 1

maxfU (); D ()g =
D() for  > N 1;
where N 1 is the point where the two graphs intersect. Clearly, N 1 is also the value
of N 1 at which the minimum of maxfU (); D ()g is achieved. Hence, the optimal
hedge-ratio satis es

U (N 1) = D (n 1) ;
22
which is equivalent to

(UN N 1)SN 1U + BNU = (DN N 1)SN 1D + BND : (3.7)


Notice that Equation (3.7) does not involve the initial endowment N 2. In particular,
N 1 is completely determined by the nal cash- ows (UN ; BNU ) and (DN ; BND ) for the
two connecting nodes at time N . Furthermore, according to Equation (2.9), we have

BN 1 = 1 +1 R f (UN N 1)SN 1U + BNU g


(3.8)
= 1 +1 R f (DN N 1)SN 1D + BND g:

Substituting the right-hand sides of (3.8) into the self- nancing equation (2.1), we see that
the strategy (N 1; BN 1) replicates the nal payo . Equations (3.7) and (3.8) show
that N 1 and BN 1 are both functions of the spot price SN 1, independent of the initial
endowment N 2.
The procedure outlined here for time tN 1 can be repeated at earlier times tN 2, tN 3,
etc. Thus, Proposition 2 follows by induction on n.
3.2 k-Convexity.
Another condition which ensures that (n; Bn ) is path-independent and replicating is
k-convexity.
De nition 3.2. The payo (N (SN ); BN (SN )) is said to be k-convex if the following
conditions are satis ed:

(i) jN 5 jN+1 for 0 5 j 5 N 1 (3.9)


and
   
(ii) 1 k SNj (jN+1 jn) 5 BNj BNj+1 5 1 + k SN (jN+1 jN ) : (3.10)
2 2

The de nition is motivated by the following important special case.


Proposition 3.3. Let F (S ) be a convex function and set
N = F 0 (SN ) and BN = F (SN ) SN F 0 (SN ): (3.11)
Then (N ; BN ) is k-convex for any k = 0.
23
Proof. According to (3.11)
j+1
SZN

jN+1 jN = F 0 (SNj+1 ) F 0 (SNj ) = F 00 (S )dS;


j
SN

and
j j j+1
ZSN ZSN SZN

BNj BNj+1 = (F (S ) SF 0(S ))0 dS = SF 00 (S )ds = SF 00 (S )ds:


j+1
SN j+1
SN j
SN

In particular, since F 00 (S ) = 0, (3.9) holds. Moreover,


j+1
SZN

BNj BNj+1 5 SNj+1 F 00 (S )ds


j
SN

= SNj+1(jN+1 jN )
 
5 1+
k S j +1
(j +1
 N)
j
2 N N

for all k = 0. Similarly,


j+1
SZN

BNj BNj+1 = SNj F 00 (S )ds


j
SN

= SNj (jN+1 jN )


 
k j +1
= 1
2 SN (N N );
j j

for all k = 0. Q.E.D.


Example 3.4 Stock options with settlement in shares have k-convex payo s. In fact, the
payo for a call option is
 
1 S =K K S=K
(S ) = and B (S ) =
0 S<K 0 S<K :

Thus, the equations in (3.11) are satis ed with


24
F (S ) = (S ) S + B(S ) = Max(S K; 0) ;
which is convex in S . Cash-settled options do not have k-convex payo s. Clearly, since
jN = 0 for all j , condition (3.10) would require BN to be constant, which is not the case
for cash-settled options.12
Proposition 3.5. Suppose that (N ; BN ) is k-convex. Then
(i) the optimal sequences fng and fBn g are path-independent, independent of the
initial endowment  1, and constitute a replicating strategy;
(ii) (n; Bn ) is k-convex for all n;
j +1
(iii) nj+1 5 nj 5 n +1 ;
(iv) the portfolios (nj ; Bnj ) satisfy the backward-induction equations

(1 + k=2) Snj+1  j +1 j  j +1
+1n+1 (1 k=2) Sn+1n+1 + Bn+1 Bn+1 ;
j j
nj =
(1 + k=2)Snj+1
+1 (1 k=2)Sn+1
j
(3.12)
  
Bnj = 1 k  j +1  j +1  j +1
1 + 2 (n+1 n )Sn+1 + Bn+1
j
1+R
  
= 1 +1 R k   
1 2 (n+1 n )Sn+1 + Bn+1 :
j j j j
(3.13)

Proof.  Assume that at time tN 1 the initial endowment is N 2. The optimal hedge
ratio N 1 is the solution to the minimization problem

min
 max
U;D
fU (); D ()g ;
with U and D as in (3.5), (3.6). Since the transcation cost function is convex, U
and D are also convex and so is

() = max fU (); D()g :


Therefore, a necessary and sucient condition for the minimum of () to be achieved
at  =  is
12
The argument shows that the only cash-settled contingent claims with k-convex payo s are bonds.
The contingent claim of Example 1.2 can be viewed as a one-period cash-settled option.
25
lim d() 5 0 and lim d() = 0: (3.14)
" d
 # d
We claim that this minimum is unique, and that it is given by the unique solution of the
problem
8
>
< U (N 1) = D (N 1)
>
(3.15)
:
 5 N 1 5 UN :
D
N

To show this, we shall prove rst prove that if the payo is k-convex then (3.15) has a
unique solution. Notice that (3.15) is equivalent to

8
>
>
1 + k2 (UN N 1)SN 1U
> 
<
+ BNU = 1 k2 (DN N 1)SN 1D + BND (3.16)
>
>
>
:
DN < N 1 < UN :
Elementary linear algebra shows that a solution to (3.16) exists if and only if
   
1 k2 (UN DN ) SN 1 D 5 BND B 5 1 + k2 (UN DN ) SN 1 U :
U
N

But this is precisely condition (3.10) of the de nition of k-convexity, so the claim follows.
It remains to show that the solution to (3.16) achieves the minimum of . For this
purpose, we verify that the rst-order optimality condition (3.15) holds. We have

dU (N 1)  k  1

k 

dN 1 = 1  2 SN 1 1 + R 1 + 2 SN 1U
   
k 1 k
5 1+
2 SN 1 1 + R 1 + 2 SN 1U
   
k U
= 1 + 2 SN 1 1 1 + R

< 0; (3.17)
and
26
dD (N 1)  k  1

k


dN 1 = 1  2 SN 1 1 + R 1 2 SN 1D
   
= 1
k S 1 1 k S D
N 1
2 1+R 2 N 1
   
k D
= 1 2 SN 1 1 1 + R

> 0; (3.18)

where we used the no-arbitrage condition D < 1 + R < U . Using (3.17)-(3.18) and the
fact that (N 1) = U (N 1) = D (N 1), we conclude that

lim d() = dU (N 1) < 0


"N 1 d dN 1
and

lim d() = dD (N 1) > 0 ;


#N 1 d dN 1
as desired.
We have shown that N 1 is the optimal hedge-ratio. Since the problem (3:16) does
not involve the endowment N 2, the optimal hedge-ratio is independent of the agent's
holdings.
The values of N 1 and BN 1 can be computed from (3.16). They are

N 1 = (1 + k=2)(1N+SNk=2)
1U (1 k=2)DN SN 1 D + B N BND
U U

SN 1U (1 k=2)SN 1D (3.19)

and
  
BN
1
1 = 1+R 1 + k2 (UN N 1)SN 1U + BNU

  
= 1 +1 R 1 k2 (DN N 1)SN 1D + BND : (3.20)

27
We now show that the optimal portfolio at time tN 1, (Nj 1; BN j 1)jN=01, is a k-convex
payo . For this, we consider the diagram
UU
N
....
.... ........
......
UN 1................
..
.... ......... ..........
... ......
.. .. ......... ......
......
... .... ......
.... ......
N 2 ......... . .. DU
.
(SN 2 =S ) ...........
.
............. N
........ ...
........ .............
........ ..
........ ............
.......
.
DN 1 .................
........
........
......
DD
N

The inequality in (3.16) implies that

UD
N 5 N 1 5 N N 5 N 1 5 N ;
and DD (3.21)
U UU D DU

and therefore

DN 1 5 UN 1: (3.22)


Using these inequalities, the formulas for Bn in (3.20) and the no-arbitrage condition
D < (1 + R) < U , we nd that

     
B 1 B 1 = 1 +1 R SUD UD k U k
N + 1 1 + 2 N 1
D U D
N N
2 N 1
   
= 1 +1 R SUD kUD
N kUN 1 + 1 + k2 (UN 1 DN 1)
 
5
1 1 + k SUD(U DN 1)
N 1
1+R 2
 
< 1 + k2 SU (UN 1 DN 1) (3.23)

28
and

BND 1 BNU 1 = 1 +1 R SUDfkUD


N kDN 1 + (1 k=2)(UN 1 DN 1)g

=
1 (1 k=2)SUD(U  D )
N 1 N 1
1+R

> (1 k=2)SD(UN 1 DN 1): (3.24)

Inequalities (3.22), (3.23) and (3.24) show that (N 1; BN 1) is a k-convex payo . We
have therefore proved Proposition 3.5 for n = N 1. The general proof now follows by
induction on n.

3.3 The backward-induction equations.


To better describe the minimum e ective cost and the optimal hedging strategy, we
introduce the state-variables13

Pn  n Sn + Bn
and
Qn  n Sn :
These variables can be interpreted, respectively, as the value of the portfolio and the value
of its equity portion at time tn.14
In the case of risk-feasible underlying assets, explicit backward-induction equations for
the pairs ( Pn ; Qn ) can be derived by solving equation (3.7).
We shall avoid the use of superscripts by using the notation

8 8
>
< PnU+1 = Un+1SnU + BnU+1 >
< PnD+1 = Dn+1SnD + BnD+1
> >
(3.25)
: :
QUn+1 = Un+1SnU ; QDn+1 = Dn+1SnD :
13
In the rest of this section, we drop the superscript  and use fn g and fBn g to represent the optimal
portfolio.
14
In this interpretation, the stock inventory is valued at the mid-price between the bid and the o er.
29
Proposition 3.6 If the underlying asset is risk-feasible, the pairs ( Pn ; Qn ) satisfy
8
>
>
>  
>
>
>
< Pn = 1+1R U PnU+1 + D PnD+1 + ( 1)a 2(1+k R) U QUn+1 + ( 1)b D QDn+1
>
>
>
>
>
>  
:
Qn = 1 PnU+1 PnD+1 + ( 1)a 2k QUn+1 + ( 1)c QDn+1 ;
(3.26)
where D ; U ; ; a ; b and c, de ned hereafter, are functions of ( Pn+1 ; Qn+1 ) (or, equiv-
alently, of (n+1; Bn+1)).
Speci cally, we have
Case 1:
Dn+1  Un+1
and
   
1 k2 SnD(Un+1 Dn+1) 5 BnD+1 BnU+1 5 1 + k2 SnU (Un+1 Dn+1) : (3.27)

In this case,
   
k
= 1+ 2 U 1 2 D ; k

U = (1 + R) (1 k=2)D ;

D = (1 + k=2)U (1 + R) ;
and
a = +1 ; b = 1 c = +1 : (3.28)

Case 2:
Un+1 < Dn+1
and

   
1 + k2 SnD(Dn+1 Un+1) 5 BnU+1 BnD+1 5 1 k2 SnU (Dn+1 Un+1): (3.29)

30
In this case,
   
= 1 k2 U 1 + k2 D ;

U = (1 + R) (1 + k=2)D ;

D = (1 k=2)U (1 + R) ;

and

a = 1 ; b = 1 ; c = +1 : (3.30)

Case 3: Either
Dn+1  Un+1
and  
1 k2 SnD(Un+1 Dn+1) > BnD+1 BnU+1;

or
Un+1 < Dn+1
and  
B U
n+1 B D
n+1 > 1 k2 SnU (Dn+1 Un+1): (3.31)
 
= 1 k2 (U D) ;

U = (1 + R) (1 k=2)D ;

D = (1 k=2)U (1 + R) ;

and
a = 1 ; b = +1 c = 1 : (3.32)

Case 4: Either
Dn+1  Un+1
31
and  
B D
B U k
> 1 + 2 SnU (Un+1 Dn+1);
n+1 n+1

or
Un+1 < Dn+1
and  
1 + k2 SnD(Dn+1 Un+1) > BnU+1 BnD+1: (3.33)

In this case,
 
k
= 1 + 2 (U D) ;

U = (1 + R) (1 + k=2)D ;

D = (1 + k=2)U (1 + R) ;

and
a = +1 ; b = +1 c = 1 : (3.34)

Proof: The proof follows by solving for N 1 in Equation (3.7) and generalizing to
arbitrary values of n.

Proposition 3.7. If the nal payo is k-convex then the pairs (Pn; Qn ) satisfy the linear
backward-induction relation

8
>
>
>  
>
>
>
< Pn = 1+1R U PnU+1 + D PnD+1 + 2(1+k R) U QUn+1 D QDn+1
>
(3.35)
>
>
>
>
>  
:
Qn = 1 PnU+1 PnD+1 + 2k QUn+1 + QDn+1 ;

with
   
= 1 + k2 U 1 k2 D ;

32
U = (1 + R) (1 k=2)D ;

and

D = (1 + k=2)U (1 + R) :

Proof: The proof is immediate, since the portfolios (n; Bn) satisfy the conditions in
(3.27) for all n.
Remark 3.8. The e ective cost of hedging a short position in a European option settled
in shares can be calculated from (3.35) for arbitrary values of k. On the other hand, a long
option position corresponds to a linear backward-induction equation de ned by Case 2 of
Proposition 3.6 | provided that the underlying asset is risk-feasible. In both cases the
corresponding values coincide with those obtained by Boyle and Vorst (1992).
Remark 3.9. In general, D and U satisfy D + U = 1. For k > 0, D and U
are positive (in all four cases) only if the underlying asset is risk-feasible. If this condition
holds, D and U can be interpreted as \risk-neutral" probabilities for no-arbitrage pricing
with transaction costs. Otherwise, since at least one of the inequalities in (3.1) does not
hold, we have D  U < 0 in Case 2 (cf. (3.30)). Thus, the scheme (3.26) becomes
numerically unstable and replicating strategies may result in negative or in nite prices as
N % 1.
If k = 0, we have

U = 1 +U R DD D = U U 1 D R
and the backward-induction equations reduces to the classical result of Cox, Ross and
Rubinstein(1982)
Pn = 1+1R (U PnU+1 + D PnD+1)
8
>
<
(3.36)
>
: PnU+1 PnD+1
Qn = U D :

33
4. k -Convex Payoffs or Risk-Feasible Underlying Assets:
Asymptotics for N % +1

This section studies the backward-induction equations for N % +1. We recall rst
the standard Cox-Ross-Rubinstein scaling of the binomial tree, in which the parameters
U; D; R, and N are related to the annualized volatility, the interest rate, and the maturity
of the contingent claim. We then show that the optimal dynamic portfolio can be approxi-
mated by the solution of the partial di erential equation (1.2)-(1.3) and its derivative with
respect to the spot price.
4.1 Adjusting the model parameters.
We denote by T the time-to-maturity of a European-style derivative security and by

dt = NT
the time-interval between successive adjustments of the hedging portfolio. Following the
classical Cox, Ross, Rubinstein (1982) scaling for the binomial model, we set
8 pdt +  dt
>
< U = e
> p +  dt
(4.1)
:
D=e  dt

and

R = erdt 1 
= rdt; dt  1: (4.2)
Here,  and r are the annualized volatility and interest rate, respectively. The parameter
 represents the subjective drift (not adjusted for risk).15
The Leland number

A  pk : (4.3)
 dt
represents the \ percentage transaction costs per standard deviation for a single period".
Lemma 4.1. For dt suciently small, the underlying asset is risk-feasible if and only if
A < 1.
15
As in the classical CRR theory, the parameter  will not appear in the asymptotic equations.

34
Proof. It follows from (3.1) that the underlying asset is risk-feasible if and only if
 
k < min U (1 + R) ; 1 + R D : (4.4)
2 U +1+R 1+R+D

Using the scaling relations (4.1), we nd that, for dt  1,


p p
U (1 + R) =  dt + O(dt)   dt
U + (1 + R) 2 + O pdt 2

and
p p
1 + R D =  dt + O(dt)   dt :
1 + R + D 2 + O pdt 2
p
Thus, for dt suciently small, risk-feasibility is equivalent to k=2 <  dt =2 , or just
A < 1. Q.E.D.
4.2. A Comparison Principle.
h i
De nition 4.2. A hedging strategy fe j gNj=n; Ben is said to be -optimal at time tn for
the payo (N ; BN ) if:
h i
(i) fe j gNj=n; Ben +  is a dominating strategy for (N ; BN );
and
h i
(ii) the e ective cost of fj gj=n; Bn di ers from the optimal e ective cost V (; Sn)
e N e
of (N ; BN ) by at most  dollars, i.e.,
 

V (; Sn ) Ben + (e n )Sn 5 :

Let us represent formally the non-linear backward-induction equations (3.26) as



Pn  =  Pn+1  n = 0; . . . ; N 1 : (4.5)
Qn Qn+1

35
Proposition 4.3. Consider a nal payo (N ; BN ). Assume that either this payo is
k-convex or that the underlying asset is risk-feasible. Suppose that there exist sequences
(Pen; Qen ) and (pn; qn ) having the following properties:
(i) e 
Pn =  Pen+1  + pn; n = 0; . . . ; N 1; (4.6)
Qen Qen+1 qn
(ii)
PeN = PN ; QeN = QN ; (4.7)
(iii) if the underlying asset is not risk-feasible, then for every n the intermediate \port-
folios" e   
n  Qen=Sn (4.8)
Ben Pen Qen
are k-convex for all n.
De ne  (N n)   
n  1 (1 + R)  nmax k
(1 + R) 1 R 0 >n;S jpn0 (S )j + 2 jqn0 (S )j :
h i
Then, for each n ; 0 5 n 5 N 1 ; the strategy f gj=n e j N ; Ben is n optimal.
Proof. Let us introduce an auxiliary contingent claim with nal payo (N ; BN ) and
with intermediate payo s (\coupons")
   
dn = qn=Sn ; (4.9)
bn pn qn
for n = 0; :::; N 1. The payo s of this auxiliary security are represented in Diagram 4.1.
Let Ven(; Sn) represent its optimal e ective cost given an endowment  at time tn .
>From the assumptions (i) through (iii), we conclude that ( e n; B
en ) is the optimal risk-
averse strategy for hedging this security (see Remark 2.2), and that (Pen; Qen ) represents
the optimal portfolio in the variables (P; Q).
The minimum e ective cost Ven (; Sn) can be estimated by stripping the intermediate
coupons from the nal payo and pricing them separately. For n0 = n, the coupon ( p0n ; qn0 )
can be purchased at
 
pn0 + k2 jqn0 j 5 nmax k
">n;S jpn" (S )j + 2 jqn" (S )j  "n (4.10)

dollars. Its value at time tn is therefore at most "n(1 + R) (n0 n). Since the e ective
cost of the auxiliary claim at time tn is at most equal to the e ective cost of the original
36
... (33 ;B33 )
.........
....
(d22 ;b22...).........
.. ..
......... ..........
... ......
... ...... ......
......
(d11 ;b.11...)......... . . ... (23 ;B32 )
.... ........... ....
... ...... (d12;b12.).........
. ........ ...... ... ..
...... ... ...
0 0 ...
(d0 ;b0 ) ......... ... ................
...... ... ......
...... ... ...... ......
......
...... ..... ......
.........
(d01;b01 ) ........... ... (13 ;B31 )
...... ... ... .
..
...... . ...
...... ....... ..
.......
(d02 ;b02 )..............
......
......
......
(03 ;B30 )

Diagram 4.1.

contingent claim plus the present value of the estimated costs of the intermediate portfolios,
we conclude that
NX1
Ven(; Sn) 5 Vn (; Sn) + "n (1 + R) (j n) = Vn(; Sn) + n : (4.11)
j =n

Similarly, a lower bound for Ven(; Sn) is obtained by considering an auxiliary derivative
security that delivers, i intermediate coupons, " dollars at each trading
h in addition to the
e N
date. The strategy fj gj=n; Bn + n is clearly a dominating strategy for such security.
e
But, since the intermediate payments are all positive on account of (4.10), the strategy is
also dominating for the European contingent claim with one single payo (N ; BN ) at
time tN . Thus, we have

Vn(; Sn) 5 Ven(; Sn) + n : (4.12)

The proof of Proposition 4.3 is complete.

37
4.3 Asymptotic analysis.
We consider rst payo s of the form
8
>
< N (SN ) = F 0 (SN )
>
(4.13)
:
BN (SN ) = F (Sn) SN F 0 (SN ) ;
where F (S ) is a four-times continuously di erentiable function satisfying
8 9
4 j =
S j d dS
<X F (S )
kF k  sup :
j ; <1: (4.14)
S j =0
The boundedness and regularity assumption (4.14) is mathematically convenient but not
essential. In Section 4.4 we extend the results derived for the payo s (4.13) to option
portfolios.
The main result of this section is:
Theorem 4.4. Let F (S ) be a function satisfying the regularity condition (4.14). Assume
that A < 1 or F(S) is convex. Let P (S; t) be the solution of the nal-value problem
8
>
> h  i
> @P + 1  2 1 + A sign @ 2 P2 S 2 @@S2 P2 + rS @P
@S rP = 0
>
>
< @t 2 @S
>
(4.15)
>
>
>
>
:
P (S; T ) = F (S ):
Set
8
< Pn (Sn ) = P (Sn; tn)
> e

>
(4.16)
: e
Qn(Sn) = Sn @P
@S (Sn ; tn ) ;
and
e   
n = Qen =Sn : (4.17)
Ben Pen Qen
Then:
(i) There exists a constant

C = C (r; ; ; A; T ) independent of F and dt, such that
Pn(Sn); Qn (Sn) satis es the backward-induction equation
e e

38
e 
Pn =  Pen+1  + pn (4.18)
Qen Qen+1 qn
for n = 0; . . . ; N 1, with

jpnj 5 C kF k(dt) 23 and jqnj 5 C kF k dt : (4.19)

(ii) In particular, consider a European-style contingent claim with payo


8
>
< N (SN ) = F 0 (SN )
>
:
BN (SN ) = F (SN ) SN F 0 (SN ) :
h i
Then, fe j gNj=n; Ben is a C 0kF k(dt)1=2 {optimal hedging strategy for (N ; BN ),
where C 0 = C 0(r; ; ; A; T ) is a constant independent of F and dt. Furthermore,
the minimum e ective cost function of this contingent claim, Vn(; Sn), satis es,
for all  and Sn,

 

Vn (; Sn ) Pen + k jQe S j S 5 C 0kF k(dt)1=2 : (4.20)
2 n n n

Proof. In view of Proposition (4.3), the proof consists in establishing that Pen  is an
Qen
approximate solution to the backward-induction equation (3.26). We shall estimate the
remainders pn and qn in (4.19) using Taylor series expansions of pn and qn in powers of
p
dt.
Step 1: Reduction to r = 0: We can assume without loss of generality that R = r = 0.
In fact, this amounts simply to measuring values in dollars-at-expiration rather than in
present value.
Step 2: P (S; t) is locally convex: If PSS (S; t) = 0 in a neighborhood of the node of
interest, then the portfolio e n+1 ; Ben+1 is locally k-convex, i.e., condition (3.27) (Case 1
of Proposition 3.6) holds.
Using the de nitions , D and U in (3.28), we nd after some calculation that
p
 (2 + A)  dt ;
39
U  12  + (1=2)2 (1 + A) pdt ;
(2 + A) 
and
D  21 +  + (1 =2)2 (1 + A) pdt : (4.21)
(2 + A) 
Expanding Pen+1 and Qen+1 in a Taylor series around (Sn ; tn+1), we nd that

PenU+1 = Pen+1 + (SnU Sn)PeS;n+1 + 21 (SnU Sn)2 PeSS;n+1 +


h i
3
O jjS PSSS jj0 dt 3= 2
p
= Pen+1 + SnPeS;n+1  dt + 21 ( Sn PeS;n+1 + Sn2 PeSS;n+1 ) 2 dt +
h i
2 3
O (jjS PSS jj0 + jjS PSSS jj0) dt 3= 2 ; (4.22)

p
PenD+1 = Pen+1 SnPeS;n+1  dt
+ 12 ( SnPeS;n+1 + Sn2 PeSS;n+1 ) 2 dt +
h i
2 3
O (jjS PSS jj0 + jjS PSSS jj0) dt3 = 2 (4.23)
and

Pen+1 Pen = Pet;n+1 dt +


h i
O kj@t P jjt;1=2 (dt)3=2 (4.24)

where we used the abbreviations PeS;n+1 = @P @P


@S (Sn ; tn+1 ), Pet;n+1 = @t (Sn ; tn+1), etc., and
the semi-norms

jjjj0  sup j(S; t)j


S;t
and
0
k kt;1=2  sup j(S; t)j + sup0 j(S;jtt) t0j1=(2S; t )j :
S;t S;t;t
40
Substituting these Taylor expansions into (3.26) (with R=0), and using the fact that
P (S; t) satis es the PDE (4.15), we nd after a tedious but straightforward calculation
that the pairs QPeenn and QPenn+1
 e 
+1
satisfy

Pen

U PenU+1 + D PenD+1
 k U QeU D D

2 n+1  Qn+1 = pn
e (4.25)

and

Qen 1 PeU PeD  k QeU + QeD  = q ; (4.26)


n+1 n+1 2 n+1 n+1 n

with
h i
jpnj S 2 PSS 3 3=2
5 C 0 + S PSSS 0 + kPt kt;1=2 (dt) (4.27)
and
h i
jqnj S 2 PSS 3
5 C 0 + S PSSS 0 + kPt kt;1=2 dt : (4.28)

Here, C = C (; ; A; T ) is a constant independent of F and dt. Hence from Proposition


A.1 in the Appendix, we conclude that

jpnj 5 C (; ; A; T ) jjF jj (dt)3=2 (4.29)


and

jqn j 5 C (; ; A; T ) jjF jj dt : (4.30)

Step 3: P (S; t) is locally concave: Next, we consider the case where PSS (S; t) < 0 in
a neighborhood of the node (Sn; tn). A calculation completely analogous to the previous
one shows that

U PenU+1 + D PenD+1 + k2 U QeUn+1 D QeDn+1 = pn


   
Pen (4.31)

and

Qen 1 PeU PeD  + k QeU + QeD  = q ; (4.32)


n+1 n+1 2 n+1 n+1 n

41
where pn and qn satisfy (4.29) and (4.30). Thus, if P (; t) is concave as a function of S in a
neighborhood of the node, the recursion relation (3.26) corresponding to Case 2 is satis ed
to the desired order of approximation.
Step 4: Local concavity and nondegeneracy imply Case 2: Notice that we have not ver-
i ed in Step 3 that the portfolios (e n+1 ; B
en+1 ) satisfy Case 2 (condition (3.29) of Proposi-
tion 3.6). This is necessary, since otherwise the estimates (4.31) and (4.32) are not relevant.
We now show that if P (S; t) is locally concave, then Case 2 indeed applies, provided that
PSS satis es a non-degeneracy condition.
Setting P (S ) = P (S; tn+1), we have, as in the proof of Proposition 3.3,
SZn U
e Dn+1 eU
n+1 = jPSS (X )j dX
Sn D

and
SZn U
BenU+1 BenD+1 = X jPSS (X )j dX :
Sn D

We estimate the latter integral as follows:

SZn U SZn U SZn U


X jPSS (X )j dX = Sn D jPSS (X )j dX + (X Sn D) jPSS (X )j dX
Sn D Sn D Sn D

SZn U
= Sn D (e Dn+1 e Un+1 ) + (X Sn D) jPSS (X )j dX
Sn D

eD
5 S n D (n+1 e Un+1 ) + 21 (max jPSS j )(Sn U Sn D)2
 
5 Sn D +
1 max jPSS j (S U Sn D) (e Dn+1 e Un+1 ) : (4.33)
2 min jPSS j n
Here, the maximum and minimum are taken over all X 2 [Sn D ; Sn U ]. Similarly,
42
SZn U  
jPSS (X )j dX = Sn U
1 max jPSS j (S U Sn D) (e Dn+1 e Un+1 ) :
2 min jPSS j n
Sn D
(4.34)
We conclude that if

D + 12 max jPSS j (U D) 5 (1 k)U (4.35)


min jPSS j 2
and

U 1 max jPSS j (U D) = (1 + k2 ) D (4.36)


2 min jPSS j
e n+1 ; B
then condition (3.29) is satis ed by ( en+1 ). Substituting the values of U; D; and
k into these estimates, we nd that both inequalities hold for dt  1 if
max jPSS j < 2 A: (4.37)
min jPSS j 2
But clearly, since dt  1,

max jPSS j  1 + jjS 3 PSSS jj0  pdt


min jPSS j min jX 2PSS (X )j
C jjF jj p
5 1 +
min jX 2PSS (X )j  dt ; (4.38)

from the a-priori estimates given in the Appendix. Thus, we conclude that if PSS (X ) is
negative in a neighborhood of (Sn tn) and

min j X 2 PSS (X )j = 2 C jjF jj pdt ; (4.39)


Sn D 5 X 5 Sn U 2 A
then the pairs (Pen ; Qen ) and (Pen+1 ; Qen+1) satisfy the backward-induction relation (4.18)
within the desired order of accuracy (4.19) guaranteed by Step 3. Inequality (4.39) can be
viewed as a non-degeneracy condition for the second derivative of P (S; t) that guarantees
that the solution of (3.26) is well-approximated locally by (Pen ; Qen ).
43
Step 5: The marginal cases: To establish that (Pen ; Qen ) is a suitable approximate solu-
tion to (3.26) in general, we now assume that (4.39) is not satis ed, and focus on Cases 3
and 4 of the backward-induction equation.16
Consider the nite-di erence equation (3.26) with parameters D , D , , a, b and c as
in Case 3. We have
A dt )2  pdt
8 p
>
>
>
 (1 2
>
>
>
A2 )  pdt
<
>
U  2+4A 1 2
2  + (1=2) (1 2 (4.40)
>
>
>
>
>  p
D  2 4A + 21  + (1=2)2 (1 A22 ) dt :
:

Substituting these formulas and the Taylor expansions of Pn+1 and Qn+1 into (3.26) leads
to

U PenU+1 + D PenD+1 + k2 U QeUn+1 D QeDn+1 = pn


   
Pen (4.41)

and

Qen 1 PeU PeD  + k QeU eD



n+1 n+1 2 n+1 n+1 = qn ;
Q (4.42)

with error terms satisfying, respectively,


2 2
pn = 2 A + A2 Sn2 PSS Sn ; tn+1 dt + O ( jjF jj (dt)3=2 ) ;

(4.43)

and
p
qn = A Sn 2 dt Sn2 PSS Sn ; tn+1 + O ( jjF jj dt ) :

(4.44)

To estimate the remainders, we use the fact that S 2 PSS (S; t) is small. In fact, since
(4.39) does not hold, we may assume that
p
j Sn2 PSS Sn ; tn+1 j < 42C jjFAjj dt :

(4.45)
16 Roughly speaking, Cases 1 and 2 correspond to those nodes where the dollar-weighted curvature
S jPSS (S; t)j is suciently large. Cases 3 and 4 correspond to the vicinity of in ection points of P (S ;t).
2

44
Substituting this inequality into (4.43) and (4.44), we conclude that pn and qn satisfy the
error estimates (4.19) when Case 3 holds. We omit the analysis of Case 4 , which is entirely
similar.
Step 6: Globally Convex Solutions: If F(S) is convex then by the maximum principle
P (S; t) is always convex. Therefore, according to Step 2 we conclude that the backward-
induction relation is satis ed within the error bounds given by (4.19). Furthermore, the
portfolios (e n; Ben) de ned by (4.8) are always k-convex according to Proposition (3.3).
Step 7: Comparison Principle: The Theorem now follow from Proposition 4.3. In
fact, applyingh this result ini conjunction with the estimates on the remainders pn and qn,
we nd that f e j gN ; B
j =n en is a  -optimal hedging strategy with

  C kF k(dt)1=2 + A
2 C k F k ( dt )1=2  C kF k(dt)1=2 ;

where C denotes a constant which depend on , , A and T , but not on dt or F . In


particular,

 

Vn (; Sn) k
Pn + 2 jQn Snj Sn = Vn(; Sn)
e e

Ben + (e n

)Sn

5 C kF k(dt)1=2 : (4.46)

The proof of Theorem 4.4 is complete.

4.4 Application to pricing and hedging option portfolios.


We extend the result of Theorem 4.4 to option portfolios.
Theorem 4.5. Consider a European-style contingent claim with a payo ((SN ); B(SN ))
with (SN ) uniformly bounded. Suppose that the function

F (SN )  (SN ) SN + B(SN )


is piecewise linear and that F 0(SN ) has nitely many discontinuities.17
Let P (S; t) be the solution of the nal-value problem (4.15) and set
17 We shall refer to this function as the payo function of the contingent claim.

45
Pn = P (Sn ; tn) ; Qn = P (Sn ; tn) Sn @P (S@Sn ; tn) :
Then, we have
 

Vn (; Sn ) k
Pn + 2 jQn Snj Sn 5 C (dt)1=8 ; (4.47)

where C depends on F and Sn but not on dt.

Proof. We shall proceed by approximations, reducing the proof to the case of the bounded
smooth payo s treated in Theorem 4.4.
Step 1. Reduction to bounded payo : Consider a derivative security with the modi ed
nal payo
8
>
< N (SN ) (0)
>
:
BN (SN ) ;
where (0) represents the (constant)
 slope of the function F at in nity. Clearly, the payo
function N (SN )   SN + BN (SN ) = F (SN ) (0)  SN is uniformly bounded.
(0)
As observed in Remark 2.3, the addition or subtraction of a xed number of shares to the
nal payo of a contingent claim induces a trivial change in the optimal hedging strategy:
the hedge-ratio is simply translated by the constant number of shares. Notice also that the
solution of the nonlinear PDE (4.15) with the modi ed nal payo F (SN ) (0)  SN is
given by P (S; t) (0)  S , where P (S; t) solves (4.15). The modi ed Delta is @P
@S  .
(0)
Thus, since the hedging problem and the PDE are both invariant under the addition of
a linear function to the nal payo we may assume without loss of generality that F is
bounded.
Step 2. Approximation by bounded smooth payo : Given any bounded, continuous
piecewise linear function F satisfying the assumptions of the theorem and any  > 0, we
can construct a function F(SN ) with the following properties:

jF(SN ) F (SN ) j 5  ; (4.48)


jF0(SN ) j 5 C ; (4.49)
where C is a constant and
46
kF k 5 3 :
C (4.50)

The function F () can be obtained by a standard convolution procedure. To derive the
estimate in (4.50) we used the fact that F 0 is uniformly bounded and that the norm k  k
incorporates derivatives of F up to third order.
De ne the auxiliary payo

(N)(SN ) = F0 (SN ) and BN()(SN ) = F(SN ) F0 (SN ) SN ;


and the auxiliary sequence of portfolios f((n); Bn())g, obtained by solving the nonlin-
ear PDE (4.15) iwith nal condition F. According to Theorem 4.4, the trading strategy
h
f(j)gNj=n; Bn() is C kFk(dt) =2 {optimal for the auxiliary payo ((N); BN()). In partic-
ular, let Vn()(; S ) denote the corresponding optimal e ective cost function. Then, from
(4.46) and (4.50) we have

()
V (; Sn )

Pn() + k2 jQ(n) Snj Sn


5
C (dt)1=2 ; (4.51)
n 2
where Pn() and Q(n) are de ned in the usual way using the solution of the PDE. From the
estimates (4.48) and (4.49) we conclude that

jP (S; t) P ()(S; t)j 5 


and

@P ()(S; t)
5 C:
@S
Using these estimates and (4.51), we nd that

 1=2

Pn + k2 jQn Snj Sn 5 C (dt3) + C (dt)1=2 ;
()
V (; Sn) (4.52)
n

where C is a constant that depends on F but not on dt.


Finally, let us compare Vn (; Sn) and Vn()(; Sn). For this purpose, consider the
situation of an agent that must deliver the payo (N ; BN ) at the expiration date but
47
holds instead the portfolio ((N); BN()). To deliver the latter payo , he must incur an
additional cost of
F (SN ) F()(SN ) + k2 j(SN ) F(0)(SN )jSN
dollars. However, since from estimate (4.48), the di erence in the payo functions is O()
and that j(SN ) F0 (SN )j < C 0 for some constant C 0, the agent can meet the nal
liability by adding to the initial portfolio (n(); Bn()p) a xed amount of shares and bonds.
Namely, he can add to his initial portfolio C 0 A  dt=2 shares, and bonds with a face
value of  dollars. The cost of this (static) portfolio is (to leading order in dt1=2) less than
p
 e r(T tn ) + A2  C 0 Sn dt :
Hence,
p
Vn()(; Sn) 5  e r(T tn ) + A2  C 0 Sn dt :


Vn (; Sn) (4.53)

Combining estimates (4.52) and (4.53), we obtain


(dt)1=2
   
Pn + k2 jQn Snj Sn

Vn (; Sn ) 5 C  + (dt)1=2 + : (4.54)
3

It is clear that the magnitude of the regularization parameter  which minimizes the right-
hand side of this last expression is  / C (dt)1=8 . The conclusion of Theorem 4.5 follows.
Remark 4.6: Portfolios of European-style calls and puts | settled in cash or in shares |
satisfy the assumptions of Theorem 4.5. The theorem also applies to barrier options with
continuous payo s. We note, however, that Theorem 4.5 does not apply to contingent
claims with discontinuous payo s, such as digital options or \up-and-out" barrier options.
Nevertheless, even in the case of discontinuous payo functions, the solution of (4.15) pro-
vides satisfactory numerical approximations to the BLPS cost function (see, for instance,
Figure 1.4 in the Introduction).
Remark 4.7 In the course of the proof we obtained also an O dt1=8 | optimal hedging
strategy, in which the hedge-ratio is the derivative of a regularized solution of the PDE
in (4.15) with payo F . This result is weaker than Theorem 4.4, since we have not
demonstrated the dt | optimality of the strategy which uses the Deltas ( @P @S ) of the
exact solution of (4.15). The reason for this is that Theorem 4.4 assumed strong regularity
condition of the payo function in order to obtain error estimates for residuals. A proof of
the optimality of the hedging strategy n = @P (@S Sn ;tn ) can be done at the expense of a
more elaborate analysis, taking into account the behavior of the Deltas in a neighborhood
48
of the jump discontinuities of F 0 , near expiration. We shall not elaborate on this point
further, observing that the dt | optimality of such strategy is well-supported by Monte-
Carlo simulations (Avellaneda and Paras (1994), for piecewise linear payo s.
We note also that Theorem 4.5 does not apply directly to contingent claims with discon-
tinuous payo functions, such as digital options. Nevertheless, payo s with discontinuities
can be super- and sub-dominated by piecewise linear functions which approximate the pay-
o function everywhere except near points of discontinuity. Even in the case of contingent
claims with discontinuous payo functions, the solution of (4.15) provides an satisfactory
numerical approximation to the BLPS cost function for A < 1 (see Figure 1.4).

49
5. Non-convex payoffs with A = 1
and Path-dependent strategies

We characterize the solution of the BLPS algorithm for contingent claims with mixed
convexity in the regime A = 1. Under these circumstances, the optimal strategies are path-
dependent and the BLPS algorithm does not correspond to the the backward-induction
equation (3.26). The optimal strategies and their e ective cost can be characterized in the
limit N % +1 by means of a nonlinear PDE corresponding to an obstacle problem for a
Black-Scholes PDE with an adjusted volatility.
5.1 The Obstacle Problem and its associated hedging strategy.

De nition 5.1. We say that P (S; t) is the solution to the obstacle problem with payo
F (S ) if
P (S; t) = e r(T t) F (Ser(T t)) ; (5.1)

@P + 1 2 (1 + A)S 2 @ 2 P + rS @P rP 5 0 (5.2)
@t 2 @S 2 @S
for all (S; t) and
@P + 1 2(1 + A)S 2 @ 2P + rS @P rP = 0
@t 2 @S 2 @S
whenever
P (S; t) > e r(T t) F (Ser(T t)) ; (5.3)
with

P (S; T ) = F (S ) : (5.4)
Remark 5.2. This free-boundary problem has a unique solution for any bounded, contin-
uous payo function F (S ) and is mathematically well-posed. It represents a natural \con-
tinuation" of (4.15) for A = 1, which corresponds formally to solving the nal-boundary
problem for a degenerate nonlinear PDE with a di usion coecient equal to 2 (1 + A) for
PSS > 0 and 0 for PSS 5 0 (in the sense of \viscosity solutions" (see Crandall, Ishii and
Lions (1992)). For
p convex payo s, it reduces to the well-known Black-Scholes equation
with volatility  1 + A proposed by Boyle and Vorst (1992) (cf. also Section 4).
The contact set associated with the obstacle problem (5.1)-(5.4) is
n o
C = (S; t) : P (S; t) = e r(T t)F (er(T t) S ) ; t < T : (5.5)
50
This is a closed set in the (S; t)-plane. We shall denote its boundary by @ C.
We will denote by Rn the ray in the (S; t)-plane which joins the points (Sn 1; tn 1) and
(Sn; tn).
De nition 5.3. We de ne the hedging strategy associated with the obstacle problem as
the strategy with initial cash reserve given by

B0 = P (S; t) 0 S (5.6)
and with a sequence of stock-holdings fng obtained through the following two phase
process, starting in Phase I if (S0; t0) 2= C and in Phase II otherwise.

Phase I: Delta-Hedging: The hedge-ratio is adjusted at every trading date to


n = @P
@S (Sn; tn ) ; (5.7)

until the rst trading date tn0 at which the ray Rn0 intersects @ C or n0 = N . If tn0 = tN
the stock-holdings are adjusted to N = F 0 (SN ) and the strategy stops. Otherwise, the
strategy goes into Phase II.
Phase II: Static Hedging: Denote by (S  ; t ) the point at which Rn0 intersects @ C (see
Figure 5.1).18 Thereafter, the hedge-ratio is maintained xed at the value

n = @P   0 
@S (S ; t ) = F (S ) ; (5.8)

until the rst time tn" at which (Sn"; tn") 2= C and




@P (S ; t )
F 0 (S  ) 5

@P (S ; t ) @P (S ; t ) ; (5.9)
@S n" n" @S n" n" @S n" 1 n" 1
or n" = N . If tn" = tN the stock-holdings are adjusted to N = F 0 (SN ) and the strategy
stops. Otherwise, the strategy goes back to Phase I.

The main result of this Section is

18 If n0 = 0 then (S  ; t ) = (S0 ; t0 ).

51
Theorem 5.4. Assume that A = 1. Let F(S) be a function satisfying the regularity con-
dition (4.14) and let P (S; t) be the solution to the obstacle problem (5.1)-(5.4). Consider
the European-style contingent claim with payo
8
>
< N (SN ) = F 0 (SN )
>
(5.10)
:
BN (SN ) = F (SN ) SN F 0 (SN ) ;
and let [fngNn=0; B0 ] be the hedging strategy associated with the obstacle problem (cf.
(5.6)-(5.9)). Then, [fngNn=0; B0] is a C kF k (dt)1=2 -optimal hedging strategy for
(N ; BN ), where C is a constant which depends on ; ; A and T .

Proof. The proof of Theorem 5.4 is done by deriving upper and lower bounds on the
e ective cost function Vn(; S ). Proposition 5.10 hereafter shows that P (S; t)  S
C kF k (dt)1=2 is a lower bound of the minimum e ective cost. Proposition 5.14 shows that
[fngNn=0; B0 + C kF k (dt)1=2 ] dominates the nal payo and hence Vn(; S ) 5 P (S; t)
 S + C kF k (dt)1=2 . The Theorem follows immediately from these two propositions.
Remark 5.5 The assumptions on the payo and on the regularity of the function F
are made for mathematical convenience. The generalization of Theorem 5.4 to payo s
corresponding to option portfolios is done as in the case A < 1 (See Paragraph 4.4).
We assume hereafter that r = 0, which amounts to measuring values in dollars-at-
expiration rather than in present value. The letter C will denote a generic constant which
depends on ; ; A and T and which may vary from one inequality to another.
5.2 The lower bound.
For the purposes of comparing various cost functions, it is convenient to consider models
in which the transaction cost parameter may vary from one node of the binomial tree to
another, i.e.

k = k (n ; j ) ; 05j5n5N : (5.11)

The following Lemma is self-evident:


Lemma 5.6. Consider the BLPS model with two transaction cost functions such that
k(1) (n ; j ) 5 k(2) (n ; j ) :
Let Vn(1)(; Sn) and Vn(2)(; Sn) represent the minimal e ective cost functions of a
derivative security under the two transaction cost regimes. Then
52
T-t

S
Figure 5.1. Schematic rendering of the binomial lattice near the contact set C. The
heavy dots represent the points in D C.

Vn(1)(; Sn) 5 Vn(2)(; Sn) :


Lemma 5.7. Let DC be the set of nodes (n; j ) in the binomial tree such that
( i) (Snj ; tn) 62 C
and
   
(ii) (Snj ; tn ) ; (Snj U; tn+1) \ C 6= ; or (Snj ; tn) ; (Snj D ; tn+1) \ C 6= ;.19
p
For any k = A  dt with A = 1, de ne the transaction cost function
8
>
< k if (n; j ) 62 DC
k(1) (n ; j ) = > p (5.12)
:
 dt if (n; j ) 2 DC :
19We use the notation [A; B ] to describe the segment connecting the points A and B in the (S; t) -plane.
Thus, DC can be viewed as the \discrete outer boundary" of C, formed of the nodes which are not in C
but which \connect" immediately to @ C.
53
Let Vn(1)(; Sn) denote the minimum e ective cost for the derivative security with pay-
o (5.10) assuming the transaction cost function (5.12). Then, there exists a constant
C = C (; ; A; T ) such that

p
Vn(1)(n 1; Sn) = F (Sn ) n 1Sn +  2 dt jn 1 j Sn C kF k (dt)1=2 (5.13)

for all n, Sn and n 1.


Proof. Given any A0 < 1, de ne k0 = A0  pdt and let Vn0 ( ; Sn) denote the optimal0
e ective cost function corresponding to the derivative security with payo (5.10) with k
as the transaction cost parameter. Then, by Lemma 5.6,

Vn(1)( ; Sn) = Vn0 ( ; Sn) :


Let P (A0)(S; t) denote the solution of the nal-value problem (4.15) with r = 0 and Leland
number A0 . From this last inequality and Theorem 4.4, we have

Vn(1)(; Sn) = Vn0 (; Sn)


0
= P (A ) (Sn ; tn )  Sn

k0 @P (A0)
 C kF k (dt)1=2
+ 2
@S (Sn; tn ) Sn

0
0
= P (A ) (Sn ; tn )  Sn + k2 jj Sn C kF k (dt)1=2 ;

0
where we used the fact that S PS(A ) is uniformly bounded. The desired estimate, (5.13),
will follow by letting A0 % 1, after we establish that

lim P (A0 ) (S; t) = F (S ) : (5.14)


A "1
0

To prove (5.14), we introduce the auxiliary function

W (S; t)  F (S )  (T t) ;
54
where  is a positive constant. We claim that W (S; t) is a sub-solution of the equation
@P + 1 2 1 + A0 sign @ 2 P
   2
@t 2 @S 2 S 2 @@SP2 = 0 (5.15)

for A0 suciently close to 1. In fact, since S 2 ddS2F2 5 kF k,

@W + 1 2 1 + A0 sign @ 2W
   2
@t 2 @S 2 S 2 @@SW2
 2F   2
1
= +2  2 1 + A 0 sign d S 2 ddSF2
dS 2

= 
1 2 [1 A0 ] kF k
2
= 0

whenever A0 > 1 kF2k2 . Assume henceforth that A0 is in this range. Since W (S; t) is a
sub-solution of (5.15) and P (A0)(S; t) is a solution and, moreover, W ( S; T ) = P (A0 ) (S; T )
= F (S ), we conclude from the Maximum Principle that P (A0)(S; t) = W (S; t) for t < T .
Therefore,
0
lim P (A )(S; t) = W (S; t) = F (S )
A0 "1
 (T t)  >0:

Letting  # 0 in this last inequality, we conclude that (5.14) holds, and the proof is
complete.

Lemma 5.8. Let tn0 5 tN be the rst hitting time of the set DC (with t0n = tN if the path
does not hit DC). Assume that the transaction cost is given by (5.12), and let Vcn0 (; Sn)
represent the minimum e ective cost of the contingent claim with payo
8
>
< n0 (Sn0 ) = F 0 (Sn0 )
>
(5.16)
:
Bn0 (Sn0 ) = F (Sn0 ) Sn0 F 0 (Sn0 ) :
Then, for any n < n0 and any initial endowment n 1,
55
Vn(1)(n 1; Sn) = Vcn0 (n 1; Sn ) C kF k(dt)1=2 : (5.17)

Proof. Suppose that an agent has an endowment of n0 1 shares at time


0
tn0 (before
rebalancing). The e ective cost of delivering the payo at time tn0 for n < N will be,
according to (5.16),

Vcn0 0 (n0 1; Sn0 ) = (F 0 (Sn0 ) n0 1 ) Sn0 + F (Sn0 ) Sn0 F 0 (Sn0 )


(1)
= F (Sn0 ) n0 1 Sn0 + k2 jF 0 (Sn0 ) n0 1 jSn0
p
= F (Sn0 ) n0 1 Sn0 +  2dt jF 0(Sn0 ) n0 1jSn0 :

Thus, by Lemma 5.7 and the boundedness of S F 0 (S ), we conclude that

Vn(1)
0 (n0 1 ; Sn0 ) = Vc0
n0 (n0 1 ; Sn0 ) C kF k (dt)1=2 : (5.18)
Inequality (5.17) then follows from Proposition 2.4. Q.E.D.

In the next Lemma, we estimate the function VcN0 (; S ).


Lemma 5.9. Let P (S; t) be the solution of the obstacle problem (5.1)-(5.4) with r = 0
and set

Pn = P (Sn; tn) ; Qn = Sn @P (@S


Sn ; tn) ; (5.19)
Let tn0 be stopping time de ned in Lemma 5.8. Consider a contingent claim which delivers
the payo
8
< n0 (Sn0 ) = Qn0 =Sn0
> e

>
(5.20)
: e
Bn0 (Sn0 ) = Pn0 Qn0 ;
when t = tn0 and denote its minimum e ective cost function under the transaction cost
function (5:12) by W
cn (; Sn ). Then

56
( i) Vcn0 (; Sn) = W
cn (; Sn ) C kF k (dt)1=2 (5.21)
for all n 5 n0 and all ; and

 
Pn + k2 jQn

(ii) W
n ( ;
c Sn)  Sn j  Sn 5 C jjF jj (dt)1=2 : (5.22)

Proof. (i) The functions P and @P


@S are uniformly continuous along the boundary of the
contact set and

P (Sn0 ; tn0 ) F (Sn0 ) 5 C kF k(dt)1=2;

@P
@S (Sn0 ; tn0 )
F 0 (Sn0 ) 5 C kF k(dt)1=2 :

This ensures that the payo (5.16) dominates (both in the number of shares and in bonds)
the payo (5.10), up to and error of order C kF k (dt)1=2 . Hence, (5.21) follows.
(ii) Since P (S; t) is convex in the complement of C, the portfolios
e n; B
( en )= (Qn =Sn; Pn Qn), with Pn and Qn given by (5.22) are k-convex as well
(1) P 
as k -convex. We know that Qn satis es the recursion relation
n


Pn  =   Pn+1  + pn n 5 n0 1;
Qn Qn+1 qn
with remainders satisfying

jpnj 5 C kF k (dt)3=2 and jqnj 5 C kF k dt ; (5.23)


The estimate (5.23) follows from Taylor expansions, as in the proof of Step 2, Theorem 4.4,
using the fact that P (S; t) satis es the PDE (5.3). Now, since k(1) = k in the complement
of DC, we have also
     
Pn = (1) Pn+1 + p^n n 5 n0 1 ; (5.24)
Qn Qn+1 q^n
where (1) denotes the backward-induction operator corresponding  to the transaction cost
function k(1). Notice the the remainders pq^^nn are identical to pqnn for n 5 n0 2.


57
To derive an estimate for the remainders pq^^nn00 11 , observe that k and k(1) di er only by

p
by (A 1)  dt along D C. Hence, the backward-induction operators  and (1) satisfy
p
k (1) k _ dt

at the nodes of D C. Therefore, since we have


     P 0 
p^n0 1 = pn0 1 +  (1) n
q^n0 1 qn0 1 Qn0 ;

and Pn0 and Qn0 pare uniformly bounded, the remainders in (5.24) for n = n0 1 can be at
most of order O( dt).
cn (; Sn ) now follows from Proposition 4.3.20
The estimate (5.22) for W

Proposition 5.10. Let Vn(; Sn) denote the minimum e ective cost function for the
contingent claim with payo (5.10), with constant k = 1.pLet P (S; t) represent the solution
of the obstacle problem with Leland number A = k=( dt). Then,

Vn( ; Sn) = P (Sn; tn)  Sn + k2 jSn PS0 (Sn; tn) Snj C kF k (dt)1=2 : (5.25)

Proof. By Lemmas 5.6 and 5.8, we have

Vn( ; Sn) = Vn(1)(; Sn) = Vcn0 ( ; Sn) C kF k (dt)1=2 : (5.26)

Using Lemma 5.9, eq. (5.21), we conclude that Vn(; Sn) satis es the lower bound (5.25).
Q.E.D.

20In order to obtain the right estimate from Proposition 4.3,  must be obtained by  rst adding
the residuals jpn j + k2 jqn j which are of order O (dt)3=2 , and then adding the O (dt)1=2 term from
jpn0 1 p^n0 1 j + k2 jqn0 1 q^n0 1 j.
58
5.3 Analysis of the associated strategy and upper bound.
In this paragraph, we prove that the strategy associated with the obstacle problem
dominates the nal payo . We shall denote by n the mid-value of the hedge portfolio,
i.e.,

n  n Sn + Bn : (5.27)

We shall also make use of the sequence

n  n Pn n = 0; 1; :::; N ; (5.28)

where Pn is de ned in (5.19). In particular, N represents the overall pro t/loss generated
by the strategy associated with the obstacle problem.

Proposition 5.11. Assume that the hedging


0
strategy goes through Phase I starting at
tn until tn0 . Then, for every j 2 [n; . . . ; n 2],

k j
2 j+1 j j Sj+1 = Pj+1
j (Sj+1 Sj ) Pj + j+1 ;

with jj+1j 5 C kF k (dt)3=2 : (5.29)

Proof. According to (5.7), j  PS (Sj ; tj ) for all


0
j 2 [n; . . . ; n0 1]. Therefore, it
follows by Taylor expansion that for all j 2 [n; . . . ; n 2]

k j j j Sj+1 = k2 jPSS (Sj ; tj ) (Sj+1 Sj )j Sj+1 + ^0j+1


2 j+1

= A2 2 Sj2 jPSS (Sj ; tj )j dt + ^j+1 ; (5.30)

with j^j+1j 5 C kF k (dt)3=2 :


59
Similarly, by Taylor expansion
Pj+1 Pj j (Sj+1 Sj ) = Pt (Sj ; tj ) dt + 12 PSS (Sj ; tj ) (Sj+1 Sj )2 + 0j+1

= Pt (Sj ; tj ) dt + 12 PSS (Sj ; tj ) Sj2 2 dt + j+1 ;


(5.31)

with jj+1j 5 C kF k (dt)3=2 :


Combining (5.30) and (5.31), we get
Pj+1 Pj j (Sj+1 Sj ) + k2 jj+1 j j Sj+1
 
1 2 2
= Pt (Sj ; tj ) + 2  (1 + A) Sj PSS (Sj ; tj ) dt + j+1 ; (5.32)

with jj+1j 5 C kF k (dt)3=2 :


Since (Sj ; tj ) 2= C, it follows from (5.3) (with r=0) that
Pt(Sj ; tj ) + 21 2 (1 + A) Sj2 PSS (Sj ; tj ) = 0 : (5.33)
Substituting (5.33) in (5.33) the result of Proposition 5.11 follows.
Proposition 5.12. Under the same assumption as in Proposition 5.11,
jn0 1 nj 5 C kF k tnt0 1 t tn (dt)1=2 : (5.34)
N 0
Proof. Expanding n0 1 and then applying (5.29) of Proposition 5.11 we obtain
0 2
nX 
n0 1 = n + j (Sj+1 Sj ) k j j j Sj+1
j =n 2 j+1

0 2
nX 0 1
nX
= n + [Pj+1 Pj ] + j+1
j =n j =n+1

0 1
nX
= n + Pn0 1 Pn + j+1 :
j =n+1
60
Therefore
0 1
nX
jn0 1 nj = j (n0 1 Pn0 1 ) (n Pn ) j 5 jj+1j : (5.35)
j =n+1

Applying inequality (5.30) to (5.35), the conclusion follows.


Proposition 5.13. Assume that the hedging strategy goes through Phase II starting at
time tn0 until time tn00 . Then,

n00 n0 = C kF k (dt)3=2 : (5.36)

Proof. Denote by (S  ; t ) the point where @ C intersects R 0 1. The proof of the propo-
n
sition is divided into several steps.
 t) 2 @ C we use the following convention
For any given pair (S;
Pt(S ; t )  lim Pt(S; t) ;
(S;t)!(S;
 t)
(S;t)2
=C

PSS (S ; t )  lim PSS (S; t) :


(S;t)!(S;
 t)
(S;t)2
=C

As shown in the Appendix, for any (S; t) 2 @ C, we have


Pt(S ; t ) = PSS (S ; t ) = 0 : (5.37)

We shall give a proof of the Proposition under the assumption that (Sn" 1; tn" 1) 2= C.
The case (Sn" 1; tn" 1) 2 C is analogous. Under this assumption, denote by S o the
value of S at which the segment joining (Sn0 ; tn" 1 ) and (Sn" 1; tn" 1) intersects @ C (see
Figure 5.2.A).
Step 1: Change in value of the hedging portfolio: The mid-value of the portfolio at time
tn" is

n" = n0 + PS (Sn0 1; tn0 1 ) (Sn0 Sn0 1) k jP (S 0 ; t 0 ) F 0 (S  )j Sn0


1
2 S n 1 n 1

+ F 0 (S ) (Sn" Sn0 ) k jP (S ; t ) F 0 (S )j Sn" : (5.38)


2 S n" n"
61
(S*,t *)
T-t (Sn-1 ,t n-1 )

(SO,t n-1 )

(Sn ,t n)

Figure 5.2.a. Rendering of the boundary of the contact set and of the points (Sn0 ; tn0 )
(Sn00 ; tn00 ) and S0, used in the proof of Proposition 5.13.
V *)
)(S-S
,t*
(S*
VS
* )+
* ,t
V (S

V(S,t n)

F(S)
V(S,t n")

S
O
S* S S n"

Figure 5.2.b. Representation of the solution of the obstacle problem and the associated
hedging strategy after the point (S; t) enters C. A temporary holding strategy is implemented
after the rst hitting time, t . The value of the portfolio will then vary along the tangent
line to the graph of F (S ). Delta-hedging resumes after the portfolio Delta (the slope of the
line) and the theoretical Delta coincide at the price level Sn00 . Resuming Delta-hedging at
that time allows the agent to remain above the intrinsic value F (S ) if the market continues
to rally. 62
By Taylor expansion of the seond and third terms, we nd that

PS (Sn0 1; tn0 1 ) = F 0 (S  ) + 12 PSS (S  ; t ) (Sn0 1 S ) + 1 ; (5.39)

jPS (Sn0 1 ; tn0 1) F 0 (S  )j = PSS (S  ; t ) (Sn0 1 S ) + 2 ; (5.40)


and, according to (5.9)
k jP (S ; t ) F 0 (S  )j Sn" = k jP (S ; t ) PS (Sn" 1; tn" 1)j Sn"
2 S n" n" 2 S n" n"

= A 2 S 2 P (S ; t ) dt +  ; (5.41)
2 n" SS n" n" 3

where
j1 (Sn0Sn0 1) + k2 2 Sn0 + 3 j 5 C kF k (dt)3=2 :
Substituting (5.39)-(5.41) into (5.38), and using (5.37) we get

n" n0 = F 0 (S  ) (Sn" Sn0 1 ) A 2 S 2 P (S ; t ) dt +  ; (5.42)


1
2 n" SS n" n" 4

where
4 = 1 (Sn0 Sn0 1) + k2 2 Sn0 + 3 :

Step 2: Change in the value function: A Taylor expansion of P (S "; t ") gives n n

P (Sn"; tn" ) = P (Sn0 1; tn0 1 ) + F 0 (S ) (Sn0 1 S  ) + Pt (S  ; t ) (tn0 1 t )


o
ZS
+ 12 PSS (S  ; t ) (Sn0 1 S  )2 + F 0 (S ) dS
S

SZ
n" 1

+ PS (S; tn" 1) dS + PS (Sn" 1; tn" 1) (Sn" Sn" 1)


So

+ Pt(Sn" 1; tn" 1) dt + 12 2 Sn2" 1 PSS (Sn" 1; tn" 1) dt + 5 ;


with
j5j 5 C kF k (dt)3=2 :
63
Therefore, using (5.37)
o
ZS
P (Sn"; tn") P (Sn0 1; tn0 1) = F 0 (S  ) (Sn0 1 S ) + F 0 (S ) dS
S

SZ
n" 1

+ PS (S; tn" 1) dS + PS (Sn" 1; tn" 1) (Sn" Sn" 1)


So

+ Pt (Sn" 1; tn" 1) dt + 12 2 Sn2" 1 PSS (Sn" 1; tn" 1) dt + 5 :


(5.43)

Step 3. Change in value of : Subtracting (5.43) from (5.42), and taking into account
that (5.33) holds at (Sn" 1; tn" 1), we have
o SZ
n" 1
ZS
n" n0 1 = [F 0 (S  ) F 0 (S )] dS + [F 0 (S  ) PS (S; tn" 1)] dS
S So

+ [F 0 (S  ) PS (Sn" 1; tn" 1)] (Sn" Sn" 1)


 
Pt (Sn" 1; tn" 1 (1 + A) 2 S 2 P (S ; t ) dt +  ;
1) +
2 n" 1 SS n" 1 n" 1 6

o SZ
n" 1
ZS
= [F 0 (S  ) F 0 (S )] dS + [F 0 (S  ) PS (S; tn" 1)] dS
S So

+ [F 0 (S  ) PS (Sn" 1; tn" 1)] (Sn" Sn" 1) + 6 ; (5.44)


where
j6j = j4 + 5 j 5 C kf k (dt)3=2 : (5.45)
Step 4. Estimating the lower bound in (5.44): Without loss of generality, assume that
S  < Sn00 , which implies that S  < S o < Sn" 1 < Sn". Since F (S ) is concave in [S ; S o]
then
So Z
[F 0 (S  ) F 0 (S )] dS = 0 : (5.46)
S

64
Also, P (S; tn" 1) is convex in the interval [S o; Sn" 1], therefore
F 0 (S  ) > PS (Sn" 1; tn" 1 ) (5.47)
or else the rst time condition in (5.9) would be violated. The convexity in the interval
plus condition (5.47) imply
SZ
n" 1

[F 0 (S  ) PS (S; tn" 1)] dS + [F 0 (S  ) PS (Sn" 1; tn" 1)] (Sn" Sn" 1) = 0:


So
(5.48)
A similar argument shows that (5.46) and (5.48) also hold when S  > S o > Sn" 1 > Sn",
while inequality (5.47) is reversed.
Substituting (5.46) and (5.48) in (5.44), and using estimate (5.45) the result of the
Proposition follows. (5.36) follows, and the proof of the Proposition is complete.
Proposition 5.14. Let [f g =0; B0 ] be the strategy associated with the obstacle prob-
N
n n
lem. Then, there exists a constant C = C (; ; A; T ), independent of F and dt, such that
[fngNn=0; B0 + C kF k(dt)1=2] is dominating for (N ; BN ).

Proof. The proof is equivalent to showing that  C kF k(dt)1=2. In order to do this,


N =
de ne the sequence of stopping times ft g as the sequence of stopping times at which
J
j j =0
the strategy moves from one Phase to the other, with 0 = 0 and J = N . Then by
Proposition 5.12
J 1]
[X2 [X2 ]
J 1
t t2j
2j+1   2j = C kF k (dt)1=2 2j+1
tN t0 = C kF k (dt)1=2 : (5.49)
j =0 j =0

We also have by Proposition 5.13 that


2
X
[J ]
2j 2j 1 = N C kF k (dt)3=2 = C kF k (dt)1=2 : (5.50)
j =1

Since 0 = 0, it follows from (5.49)-(5.50) that


J
X
N = j  j 1 = C kF k (dt)1=2 ;
j =1

and so the proof is complete.


65
Remark 5.15 If A = 1 and the payo has mixed convexity, the optimal hedging strategy
is not unique, as we now show. The strategy associated with the obstacle problem is such
that the agent ceases to transact as soon as the path (Sn; tn) reaches the contact set and
resumes transactions only when (i) the path exits the contact set and (ii) the number of
shares in the portfolio is near the nominal Delta (cf. equation (5.9)). A moment of thought
reveals that, since the portfolio value exceeds F (S ) as soon as the path passes through C
(assuming r = 0), the agent could choose to continue making stock transactions, as long
as the net portfolio value is constrained to remain above the obstacle. An extreme case
along these lines would be to transact the maximum amount of shares (long or short) that
would still leave the agent with a portfolio value above the obstacle at the next period.
We believe that the latter prescription corresponds to the strategies originally proposed
in Bensaid et al.(1992). Di erent dynamic strategies can be implemented according to
the agent's preference for allocating the excess returns arising from super-replicating the
payo .

References

D.G. Aronson (1967), \Bounds for the fundamental solution of a parabolic equation", Bull.
Amer. Math. Soc.,73, 890-896.
C. Albanese and S. Tompadis (1995), \Transaction Costs and Non-Markovian Delta-
Hedging", preprint, Department of Mathematics, University of Toronto.
M. Avellaneda and A. Paras (1994), \Dynamic hedging for derivative securities in the
presence of large transaction costs", Applied Mathematical Finance, 1, 165-193.
B. Bensaid, J. Lesne, H. Pages and J. Scheinkman (1992), \Derivative asset pricing with
transaction costs", Mathematical Finance, 2 (2), 63-86.
P.P. Boyle and T. Vorst (1992), \Option replication in discrete time with transaction
costs", J. Finance, 47, 271-293.
G. Constantinides (1979), \Multiperiod consumption and investment behavior with convex
transaction costs", Management. Sci., 25, 1127-1137.
G. Constantinides (1986), \Capital Market Equilibrium with transaction costs", J. Political
Econ. 94, 842-862.
J. C. Cox, S. Ross and M. Rubinstein (1979), \Options pricing: a simpli ed approach",
Journal of Financial Economics 7, 229-263
M. G. Crandall, H. Ishii and P.L. Lions (1992), \User's guide to viscosity solutions of
second-order partial di erential equations", Bulletin American Math. Soc., 27, (1), 1- 67.
J. Cvitanic and I. Karatzas (1995), Contingent claim pricing with transaction costs,
preprint, Department of Mathematics and Statistics, Columbia University.
66
M. H. Davis, V. G. Panas and T. Zariphopolou (1993), \European option pricing with
transaction costs ", SIAM J. Control and Optimization, 31, 470-493.
B. Flesaker and L. P. Houghton (1994), \Contingent claim replication in continuous time
in the presence of transaction costs", Merrill Lynch working paper.
A. Friedman (1964), Partial di erential Equations of Parabolic type, Prentice-Hall, Inc.,
Englewood Cli s, N.J.
A. Friedman (1982), Variational principles and free-boundary problems, Wiley Interscience,
New York.
E. R. Grannan and G. H. Swindle (1994), \Minimizing transaction costs of option hedging
strategies", preprint, to appear in Mathematical Finance.
S. D. Hodges and L. Clelow (1993), \ Optimal Delta-hedging under transaction costs",
preprint, Financial Options Research Center, University of Warwick.
S. D. Hodges and L. Clelow (1994), \ Gamma-hedging in incomplete markets under trans-
action costs", preprint, Financial Options Research Center, University of Warwick.
S.D. Hodges and A. Neuberger (1989), \Optimal replication of contingent claims under
transaction costs", The Review of Futures markets, 8, 222-239.
T. Hoggard, E. Whalley and P. Wilmott (1993), \Hedging option portfolios in the presence
of transaction costs", to appear in Advances in Futures and Options Research
N. V. Krylov (1985), Nonlinear elliptic and parabolic equations of second order, Mathe-
matics and its Applications
D. Kinderlehrer and G. Stampacchia (1980), An introduction to variational inequalities
and their applications, Academic Press, New York.
H. E. Leland (1985), \Option pricing and replication with transaction costs", J. Finance
40, 1283-1301
J. Nash (1958), \Continuity of solutions of parabolic and elliptic equations", Amer. J.
Math., 80, 951-954.
E. Whalley and P. Wilmott (1993), \Counting the costs ", RISK, 6 p. 10

67
Appendix: Regularity of Solutions
of the Nonlinear Diffusion Equations and
Obstacle Problems

The proofs of Theorems 4.4 and 5.4 required the regularity of the solutions of the
nonlinear PDE problems (4.15) and (5.1)-(5.4). The main result that was used is
Proposition A.1. (i) Let F satisfy (4.14) and let P (S; t) be the solution of the nonlinear
problem (4.15). Then
sup S @@SPj 5 C kF k
j
j

S;t

for 0 5 j 5 3 and

kPt kt;1=2 5 C kF k ;
where
0
kPt kt;1=2  sup jPt(S; t)j + sup0 jPt(S;jtt) t0 jP1=t2(S; t )j
S;t S;t;t

is the Holder norm of order 1=2 in t.


(ii) The same estimates hold for the solution of the obstacle problem for (S; t) in the
complement of the contact set C.

We give only an outline of the proof, and refer the reader to Friedman (1964, 1982)
and Kinderlehrer-Stampacchia (1980) for for the general regularity theory of parabolic
equations.
A.1 The case A < 1.
We shall assume without loss of generality that r = 0. This changes the PDE in (4.15)
to the simpler form

Pt + 12 ~ 2 S 2 [ PSS ] PSS = 0 ; (A.1)


with

~2 [ PSS ] = 2 [1 + A sign (PSS ) ] :


We de ne a new independent variable, y = ln S . It is straightforward to verify that
68
S PS = Py (A.2.a)

S 2 PSS = Pyy Py (A.2.b)


and

S 3 PSSS = Pyyy 3Pyy + 2Py : (A.2.c)


In particular, equation (A.1) can be rewritten in the form

Pt + 21 ~2 [ Pyy Py ]  (Pyy Py ) = 0 (A.3)


with y 2 ( 1; +1) and t 2 (0; T ). The desired estimates for the solution of (4.15)
follow immediately from

Proposition A 2. Let F (y) be a smooth function satisfying


j
sup ddyF (y; t) < 1 :
4
X
kF k 
j
(A.4)
y;t
j =0

Then,
j
sup @@yP (y; t) < C kF k ;
3
X

j
(A.5)
y;t
j =0

and
kPt kt;1=2 < C kF k ; (A.6)
where C is a constant that depends only on ; A and T .

Proof: (i) The uniform boundedness of P follows immediately from the Maximum Prin-
ciple applied to equation (A.3).
(ii) The estimate for Py follows by di erentiating formally (A.3) with respect to y.
Accordingly, setting Q  Py , we obtain

Qt + 21 @y ~2  (Qy
 
Q) = 0 : (A.7)
69
The Nash estimates for the fundamental solutions of parabolic equations with bounded,
measurable coecients (see Nash (1958), Aronson (1967)) imply that there exists a con-
stant C , which depends only on  A and T , such that

sup jQ(y; t)j 5 C sup jQ(y; T )j :


y;t 5 T y

(iii) To estimate the second derivative with respect to y, we rewrite equation (A.7) in
the form
Qt + 21 @y G (Qy Q ) = 0 ; (A.8)

where G(X ) = 2 X + 2 A jX j. Notice that G is a uniformly Lipschitz continuous


function. Set Q(h) (y; t)  Q(y + h; t). Then, from (A.8) we have

Qt(h) Qt = 1 @ G Q(h) Q(h)




2 y y

+ 12 @y G (Qy Q)

= 1@ n
H

Q (h)
Qy Q (h)
+Q
o
: (A.9)
2 y y

where H is a a function that satis es 2 (1 A) 5 H 5 2 (1 + A) (H is the derivative of


G evaluated an intermediate point). Next, de ne the di erence quotient

D(h) = h1 Q(h)
 
Q :

Equation (A.9) implies that D(h) satis es the equation

Dt(h) + 12 @y H  Dy(h) D(h) = 0 ;


h  i
(A.10)

which is structurally analogous to (A.7). In particular, we conclude that


D(h) (y; t)j 5 F h FS ;


(h)

S


0

for all h. Letting h ! 0 in this last estimate, we conclude that


70
Pyy (y; t) = hlim
!0
D(h) (y; t)
is uniformly bounded.
(iv) Next, we show that Pyyy is uniformly bounded. Setting D = Pyy and di erentiating
both sides of (A.8) with respect to y, we obtain

Dt + 12 @yy
2
fG (D Q) g = 0 : (A.11)

Subtracting equation (A.8) from (A.11), we conclude that, if  = D Q, then

@t  + 21 @yy
2
G ( )
1 @ G () = 0:
2 y (A.12)
We analyze the di erence-quotients for , namely

(h)(y; t)  h1 [ (y + h; t) (y; t)] :


It is readily seen, using the intermediate-value theorem, and equation (A.12), that these
di erence-quotients satisfy the PDEs
1 h
t + 2 @yy H  
i
1 h i

2 @y H   = 0: (A.13)
(h) 2 (h) (h)

where H is de ned as before. We can regard this last equation as the adjoint of

@t  = 21 H  @yy2  + 1 H@  ;
2 y (A.14)

which is the Fokker-Plank equation of an L1-bounded semigroup (see Krylov (1985)). This
implies that the adjoint semigroup is bounded on L1.21 Hence, the di erence-quotients
(h) are bounded uniformly in h in terms of the initial data. Passing to the limit as h ! 0,
we conclude that Pyyy is uniformly bounded in terms of kF k.22
(v) The Holder-1/2 estimate for Pt. Since jS 2 PSS j is uniformly bounded, we conclude
from (A.1) that Pt is bounded in terms of kF k.
21 L1 and L1 denote, respectively, the spaces of Lebesgue integrable and uniformly bounded functions.
22 The third derivative of P is not continuous. This can be checked by writing equation (A.11) in
divergence form (as in (A.10). This equation implies that the \ uxes" H  (Dy D) must be continuous
across the curves where PSS vanishes. Expressing this in terms of the variable S , we nd that the ratio of
+ =P
the third derivatives along such curves is given by PSSS SSS = (1 A)=(1+ A), where the superscripts
 correspond to the sign of the second derivative on each side of of the interface.
71
Let us analyze the modulus of continuity of Pt in the variable t. Di erentiating formally
equation (A.3) with respect to t and setting   Pt , we nd that

t + 21 ~2 [ Pyy Py ]  (yy y ) = 0 : (A.15)


The boundary condition for  is

(y; T ) = 1 ~2 [ F F 1 G (F
y ] ( Fyy F y) = Fy ) ;
2 yy
2 yy

which is a Lipschitz function in view of the regularity of F . The Holder-1/2 continuity of


U (y; t) in t follows from standard theory.

A.2 The obstacle problem for A = 1: Analogy with the Stefan problem of heat
conduction.

The obstacle problem (5.1)-(5.4) has many features in common with classical free-
boundary problems discussed in the PDE literature (Friedman (1964, 1982), Kinderlehrer
(1980)). To derive the regularity properties of solutions, we shall rst analyze the com-
patibility conditions satis ed by P along the free boundary. As usual, we assume that
r = 0.
Let us postulate, heuristically, that the free boundary can be represented locally as a
graph, i.e. that

(S; t) 2 @ C () S = S (t) ;
where S (t) is a smooth function. With this notation, we have

P (S (t); t) = F (S (t)) (A.16)


and

PS (S (t); t) = FS (S (t); t) : (A.17)


Here, the left-hand side is understood as the limit of the partial derivative of P as S
approaches the obstacle with (S; t) 2= C. The same convention will be used for higher-order
derivatives evaluated along @ C.
Di erentiating formally (A.16) with respect to t and using (A.17), we obtain
72
PS (S (t); t)  S_ (t) + Pt(S (t); t) = FS (S (t))  S_ (t) ;
where  denotes the time derivative, and thus
Pt (S (t); t) = 0 : (A.18)

Using the equation satis ed by P in the complement of C,

Pt + 12 2 (1 + A) S 2 PSS = 0; (A.19)

we conclude also that

PSS (S (t); t) = 0 : (A.20)


Another interesting relation is obtained by di erentiating (A.17) with respect to t.
Accordingly, we obtain

PSS (S (t); t)  S_ (t) + PS t (S (t); t) = FSS (S (t); t)  S_ (t) ;


or, from (A.19) and (A.20),

PS t (S (t); t) = FSS (S (t); t)  S_ (t) : (A.21)


Introduce as before   @t P , we nd that

t + 12 2 (1 + A) S 2 SS = 0; for  < 0 ; (A.22)

and
S_ (t) = 1
FSS (S (t)) S (S (t); t) ; (A.23)

where S = S (t) denotes the boundary of the set where  vanishes. This problem is known
as the Stefan problem.23
Notice that the speed of the free-boundary is determined by the \Stefan constant"
23 Classically, the Stefan problem describes the evolution of the temperature pro le of a two-phase
medium, such as ice and water. The interface condition (A.23) expresses that the velocity of the phase-
boundary is proportional to the temperature ux across it.
73
K (S (t)) = 1
FSS (S (t)) ; (A.24)

which depends on the curvature of the payo function F . Herein lies an important di er-
ence with regards to the classical Stefan problem. If FSS (S (t)) = 1, then the free-
boundary will move with zero speed. Also, if
FSS (S (t)) = 0 , it moves with in nite speed, which means that the function P (S; t)
at such times can be regarded as the new data for a Stefan problem with a di erent
location of the free-surface.
The two extreme cases, far from being pathological, correspond in fact the case of option
portfolios, when the payo F is piecewise-linear. The solution of the obstacle problem for
these cases was completely described in Avellaneda and Paras (1994). Due to the in nite
curvature, the free boundaries are \pinned" exactly at those points S  where the payo
function has a discontinuous second derivative and is locally concave, i.e. FSS / (S
S ). The solution of the obstacle problem reduces then to \pasting together" solutions
of two-point boundary-value problems for equation (A.19). The regularity properties of
solutions follow immediately. Aside from a nal time-interval near the expiration date T ,
where the payo function F may have concave kinks outside C, the solution of the obstacle
problem for piecewise-linear payo s satis es the estimates

S @@SPj < 1
3
X j
sup j
 > 0
j =0 (S;t)2
= C ;t = 

as well as a Holder-1/2 estimate for Pt in time for t = .


If FSS is smooth and locally FSS 6= 0, the obstacle problem has a similar structure. It
can be mapped to a sequence of classical Stefan-type problems, with nite \heat constant"
K . The regularity properties for P (S; t) stated in Proposition A.1 can then be deduced
from classical results for the Stefan problem.

74

You might also like