Spanning and Derivative-Security Valuation

Gurdip Bakshi and Dilip Madan

First Draft: May 1997

Current Version: January 1999

Forthcoming in the Journal of Financial Economics

 Bakshi and Madan are both at Department of Finance, Robert H. Smith School of Business, University of Mary-
land, College Park, MD 20742. Bakshi can be reached at Tel: 301-405-2261, Email: gbakshi@rhsmith.umd.edu, Web-
site: www.rhsmith.umd.edu/ nance/gbakshi/; and Madan at Tel: 301-405{2127, Email: dbm@rhsmith.umd.edu
and Website: www.bmgt.umd.edu/dmadan/. For helpful comments and discussions, we would like to thank Kerry
Back, David Bates, Peter Carr, Alex David, Steve Figlewski, Mark Fisher, Helyette Geman, Rick Green, Steve
Heston, Nengjiu Ju, Nikunj Kapadia, Hossein Kazemi, Inanc Kirgiz, Eduardo Schwartz, Bill Schwert, Louis Scott,
Lemma Senbet, Marti Subrahmanyam, Alex Triantis, Haluk Unal, Zvi Wiener, Xiaoling Zhang, Jianwei Zhu, and
participants at the 1999 AFA meetings in New York. The authors are especially grateful to an anonymous referee for
his help in this project. The May 1997 version of the paper was circulated under the title \A Simpli ed Approach
to the Valuation of Options." Any remaining errors are ours alone.

0
 Spanning and Derivative-Security Valuation

Abstract
This paper proposes a methodology for the valuation of contingent securities. In particular,
it establishes how the characteristic function (of the future uncertainty) is basis augmenting
and spans the payo universe of most, if not all, derivative assets. In one speci c application,
from the characteristic function of the state-price density, it is possible to analytically price
options on any arbitrary transformation of the underlying uncertainty. By di erentiating (or
translating) the characteristic function, limitless pricing and/or spanning opportunities can be
designed. As made lucid via example contingent claims, by exploiting the unifying spanning
concept, the valuation approach a ords substantial analytical tractability. The strength and
versatility of the methodology is inherent when valuing (1) Average-interest options; (2) Corre-
lation options; and (3) Discretely-monitored knock-out options. For each option-like security,
the characteristic function is strikingly simple (although the corresponding density is unman-
ageable/indeterminate). This article provides the economic foundations for valuing derivative
securities.

JEL Classi cation Numbers: G10, G12, G13

Keywords: spanning, characteristic functions, state-price density, pricing of contingent

claims, option valuation, Arrow{Debreu securities, average{rate options, correlation options,
knock-out options.

1
 The notion that options complete markets is at the core of modern nancial economics (see
the pioneering Ross (1976)). Despite its theoretical attractiveness, the idea about expanding the
asset space via European options has, however, remained mostly an abstraction. With a few
exceptions, it has not resulted in any valuation simpli cations. Clearly, the set of applications
where the option price has been exploited, directly or indirectly, to value other contingent claims
(in its basis) is potentially sparse (e.g., the pricing of elementary securities). One reconciling
possibility is that, although options span other securities, they are complex to value at the outset.
For a general stochastic structure, the diculty stems primarily from the lack of analyticity of
the option payo |a feature that has hampered closed-form option pricing characterizations. For
instance, outside of the canonical log-normal asset pricing or the Bessel interest rate class, the
fundamental valuation equation for the option price is mostly overwhelming. Even when options
are non-redundant securities and the option price is analytical, derivative{security pricing is still
not so tractable to closed-form formulations: the positioning in the continuum of options is a priori
inexplicit. Confronted with such issues, the objective of this paper is to introduce a spanning entity
with the ability to overcome the aforementioned valuation barriers. Speci cally, it is shown that the
future uncertainty's characteristic function is indeed gifted with the qualities that one should look
in a desirable spanning engine. First, its valuation is substantially more amenable (than options)
to analytical constructions. Second, the underlying basis is analytical and orthonormal. Third, it
jointly and simultaneously induces closed-form representation of every contingent claim (options
inclusive) covered by its span. As reliance is on fundamental properties of characteristic functions,
their validity is independent of how the remaining uncertainty is visualized and regardless of the
sources of randomness.
To understand the intuition behind each of the statements, it is worthwhile to pay attention
on what the characteristic function is really composed of. From a valuation standpoint, the entity
represents the price of a security that promises the holder a trigonometric payo contingent on the
remaining uncertainty. From a mathematical and economic viewpoint, it is the Fourier transform
of the state-price density function (the product of the risk-neutral density and the discounting
factor). As is well{acknowledged from Fourier theory, the characteristic functions (for a vast class)
are in nitely di erentiable from which they also inherit their smoothness and hence valuation
tractability. Needless to say, for most valuation problems that economists consider pragmatic
and interesting, the valuation equations for characteristic functions are remarkably simple, even
though the counterpart ones for state-price density, or the option price, are twisted. Actually,
one can count on centuries-old probability theory to arrive at the characteristic functions for a
comprehensive class of stochastic processes. In one extreme, a large cohort of pure-jump price
processes are recognized and mathematically represented through their characteristic functions.

1
 Strictly speaking, the span via options and the span via characteristic functions are completely
interchangeable (subject to some regularity conditions). Granted, the characteristic function is
recoverable from options; and the reverse holds as well, they are competitors for describing their
span of claims. But in light of the above discussion, that is only true in theory! Nevertheless, since
the payo on characteristic functions is separable into trigonometric sine and cosine, it has the
added distinctive trait that pricing and/or spanning can be achieved through its di erentiation or
translation (as many times as one would like). The superiority of the characteristic function as a
primary set of spanning securities is also apparent as polynomial basis can be generated from dif-
ferentiation; whereas, the opposite direction is delicate and involves summations of in nite series.
Also as one might anticipate, the positioning in the continuum of characteristic functions can be
designed scienti cally by drawing on inverse Fourier theory. Thus, so long as the characteristic
function of the state-price density is readily computable, the valuation of any arbitrary claim can
be internally accomplished.
From a practical perspective, the observation that a generic derivative-security pricing problem
is equivalent to solving just for the characteristic function is promising, and potentially vital. To
pursue the above central theme and to gauge the associated simpli cation more rigorously, we
adopted the topic of option valuation for a benchmark analysis.1 In lling this vacuum, a key
security decomposition is rst established: the traditional European call can be unbundled into its
primitives (i) the discount bond price; (ii) the scaled-forward price (the fair price of a commitment
to deliver the underlying asset at expiration); and (iii) two Arrow-Debreu securities (or, delta
claims). Unique to our treatment, however, each primitive security is spanned by the payo
on characteristic functions and can hence be valued recursively through its manipulation. More
signi cantly, options written on arbitrary (smooth) function of the underlying uncertainty can be
priced from the same rudimentary building-block: the characteristic function of the state-price
density. Particularly, this and the original valuation task are rendered feasible without conjecturing
(ad-hoc) solutions to the fundamental valuation equation of each call (provided the characteristic
function of the state-price density can be determined by solving the conditional expectation or
the corresponding valuation equation). Our examination of the problem also yields the following
insight: When the option claim is on the exponential of the uncertainty (as in equity or bond option
models), the characteristic functions corresponding to each Arrow-Debreu security are translates
of one another. When the option is on the level of uncertainty, the characteristic function for the
1 During the writing of this paper, Due, Pan, and Singleton (1998) have extended one of our option pricing
results (in an earlier version) to the ane Markov jump-di usion class. See the discussion in Section 2.1 and in the
proof of their Proposition 2 and 3. In what respects our present work is di erent from theirs will become apparent
from Theorem 1 and the explanation thereafter. For one, we make precise what collection of securities are spanned
by characteristic functions and then exploit this insight to value all contingent claims and not just options.

2
 rst (second) Arrow-Debreu security is constructed from the di erentiation (translation) of the
primitive characteristic function. In all such option problems where the two characteristic functions
are in di erent parametric classes, the embedded Arrow-Debreu securities are heterogeneous in
their probability compositions as a rule.2
To expound on the ner aspects of the approach, we consider the explicit pricing of (a) average-
rate interest rate options; (b) correlation options; and (c) discretely-monitored knock-out options.
In each application, the characteristic function of the respective uncertainty is instrumental in
spanning/closed-form pricing. In transitioning to average-interest rate options, we assume that the
short interest rate is governed by a square-root process (i.e., Cox, Ingersoll, and Ross (1985), hereby
CIR). Our inquiry imparts quite a few basic insights concerning average rate claims: (1) The call
price is (average) scaled forward price multiplied by a delta claim minus the product of the discount
bond and the second delta claim multiplied by the adjusted strike price. To obtain the adjusted
strike price, one must deduct the past average interest from the contractual exercise price; (2) The
density of the remaining uncertainty (the continuous-sum) has no analytical representation but
its characteristic function possesses an easy-to-interpret exponential-ane structure; and (3) The
characteristic function for the rst delta claim is obtained from di erentiation, while the second
by translation, of the original characteristic function. Our inspection also uncovers the nding
that the second delta security is non-central chi-squared distributed, while its twin-counterpart
does not share the same parent distribution.
Our innovations can also be applied to options written on more than one asset and especially
outside of the log-normal environment. Generating the analytical solution to the joint character-
istic function of the two assets lies at the crux of valuation (it will span all claims contingent on
the joint uncertainty). In a distinctive example of our own, we stipulate (i) a payo structure
in which the call is exercised only when the (gross) return on each asset exceeds a pre-speci ed
threshold (i.e., calls on correlation); and (ii) each asset innovation is cross-correlated and possess
a common volatility factor. Naturally, the call (put) option is in-the-money when the returns are
positively correlated in a rising (declining) market. As articulated in Zhang (1998), correlation
2 Characteristic functions have been used for claims pricing by a number of authors, starting with Heston (1993).
However, our reasons for reexamining characteristic functions are somewhat di erent from his. First, Heston's
goal is to solve a particularly parameterized stochastic volatility option model and to analytically determine the
characteristic function for each Arrow-Debreu security. Our main object of interest, in contrast, is the characteristic
function of the state-price density. Second, this study elaborates how a generic payo can be spanned by either a
continuum of characteristic functions, or by a continuum of calls. Third, by integrating the spanning and pricing
properties of the characteristic function, we put on a rm footing how it is that a large class of payo functions
can be built and valued (with or without univariate/multivariate exercise regions). By drawing on this distinction
and then breaking up a call into its pricing components, extant valuation steps can be circumvented altogether (i.e.,
the complex practice of solving each Arrow-Debreu security separately and then guessing their solution). Other
similarities/di erences with Heston and existing work will be reviewed later.

3
derivatives are desirable for coping cross-market, or cross-currency (commodity) risks. In the
context of equity markets, they even allow investors to position on a stock/sector relative to a
market index. These contracts are precisely what Ross (1976) and Nachman (1988) have labeled
complex options and joint simple options, respectively. In any case, as the composite payo is a
product of two calls, its solution structure requires four delta securities in analytical-form. Each
security may be interpreted as the expectation of a unity payo conditional on both calls expiring
in-the-money. As this is an option on the exponential of the joint uncertainty, each characteris-
tic function is translated from the joint characteristic function (and thus in the same parametric
class). In recovering each Arrow-Debreu security price, we adapt a result from Shephard (1991)
and extend the one-dimensional Fourier inversion formulation (i.e., Heston (1993) and Kendall and
Stuart (1977)) to a multi-dimensional setting. But this development cannot be adopted verbatim
and depends on the valuation problem at hand, and on the exercise region of the calls. This style
of reasoning is evident in the valuation of discretely-monitored knock-out options.
This article is organized as follows. How characteristic functions facilitate in the spanning and
pricing of contingent securities is made exact in Section 1. Section 2 is devoted to the pricing
of average-rate interest rate options. Section 3 re nes the methodology to cover claims written
on more than a single asset. A pricing formula for discretely-monitored knock-out options (N-
dimensional generalization) is proposed in Section 4. Concluding remarks are o ered in Section
5. The proof of each result can be found in the Appendix.

1 Spanning and Pricing Via Characteristic Functions

To go directly to the center of the derivative-security pricing problem and to their spanning
underpinnings, consider a generic European{style call option contract with expiration date t +  ,
strike price K , and claim payo :

max (0; X [s(t +  ); r(t +  ); y (t +  )] K ) (1)

where, for completeness of analysis, the payo on the call is contingent on the price of a traded asset
s(t), the spot interest rate r(t), and the vector of state variables y(t). To suppress unnecessary
notation, write X [s(t +  ); r(t +  ); y (t +  )] as X (t +  ) and de ne the exercise region of the call
as: X  fX (t +  ) > K g. Let X (t) > 0 with probability 1 for all t and  fX (t +  ) > 0g.
Provided certain regularity conditions are satis ed, the time t price of the option contract, denoted

4
C (t;  ; K ), is:
  Z t+  
C (t;  ; K ) = EtQ exp r(u) du max (0; X (t +  ) K ) (2)
Z  Z t+t  
= exp r(u) du [X (t +  ) K ] q( ) d; (3)
X t

where EtQf:g represents time{t conditional expectation under the equivalent martingale measure
(which is presumed to exist) and q ( ) is the risk-neutral
 (joint) density function of the remain-
R t + 
ing/future uncertainty:   t r(u) du; X (t +  ) . From Breeden and Litzenberger (1978), the
 R t+ 
state-price density is simply: q ( ) exp t r ( u ) du . Although the basic valuation problem
outlined in (2)-(3) is well-known, an often posed question is: How can the conditional expectation
and hence the derivative-security price be determined analytically? Clearly, when the density func-
tion, q ( ) (or the state-price density), is known, and tractable, the valuation problem warrants no
further simpli cation. But, unfortunately, for most realistic option pricing and derivative-security
valuation applications, the exercise region of the call/put is contingent on a general (vector)
Markov (or non-Markov) process for which the state-price density is either unknown or cannot
be expressed in terms of special functions of mathematics. As will be validated shortly, the char-
acteristic function of the state-price density is remarkably uncomplicated (in a relative sense) for
option problems of practical interest, even though the state-price density function is not. For the
most part, and as we show, all that is required for option and derivative-security valuation is the
closed-form formulation of the characteristic function.
As our simpli cations are about exploiting the fundamental properties of characteristic func-
tions and their span, the principal approach will be applicable regardless of the source of primitive
uncertainty: whether in discrete-time, continuous-time, pure-jump, or mixture environments. Let
x(t)  (s(t); r(t); y(t)0)0. Since option and claim valuation problems are conventionally cast in a
di usion setting, for now, assume that x(t) is a vector Markov Ito process as displayed below:

dx(t) = [x(t); t] dt + (t); t] d!(t); (4)

where ! (t) represents a (vector) standard Brownian motion. Under the appropriate set of regu-
larity conditions and the dynamics (4), the solution to the valuation partial di erential equation
of the call below:
1 tr  0C  +  C C r C = 0 (5)
2 xx x 
subject to C (t + ; 0; K ) = max (0; X (t +  ) K ) is, from the Feynman-Kac theorem, the con-
ditional expectation (2). Leave the exact dynamics for [x(t); t] and  [x(t); t] unspeci ed for the

5
moment.
Traditionally, researchers have directly solved contingent claims valuation equations such as
in (5) (e.g., see the seminal work by Cox and Ross (1976); Cox, Ingersoll, and Ross (1985); and
Merton (1973) on this topic). But, is that necessary or ideal for a general claims problem? Can we
somehow algebraically span the underlying payo and then price the claim? What are the distinct
advantages to adopting one approach over the other? To render these statements more precise
and to seek answers to the above questions, de ne the characteristic function of the state-price
density as (see the classic Lukacs (1960)):
  Z t+  
f (t;  ; )  EtQ exp r(u) du  e i  X (t +  ) (6)
Z t  Z 
t+
= ei  X (t+ ) exp r(u) du q( ) d; (7)
t

which is implicitly the time-t price of a hypothetical claim that pays ei X (t+ ) (where i = 1
p
and  is some parameter of the contract) at date t +  . Since ei X (t+ ) = cos (X (t +  )) +
i sin (X (t +  )) by Euler's identity, the payo on characteristic functions is mathematically com-
posed of trigonometric sine and cosine.3 Ignoring extreme counterexamples, the characteristic
function is in nitely di erentiable
Z  Z t+ 
exp r(u) du (iX )n ei  X q( ) d < 1 n = 1; 2;


; 1
t
with nite algebraic moments of all orders. The characteristic function satis es
1 tr  0 f  +  f f r f = 0 (8)
2 xx x 

subject to f (t + ; 0; ) = eiX (t+ ). While the valuation equation (8) and its surrogate in (5)
are observationally indistinguishable, the boundary condition for the characteristic function is
3 Technically, the characteristic function formulated in (7) is well-de ned even without the inclusion of time-value
factor. Indeed every admissible characteristic function in the classical theory is unity at  = 0. Certainly, what we
have described in (6) is the intrinsic value of a trigonometric payo . It is therefore not improper to call f (t;  ; )
a discounted (or, spot) characteristic function. Subject to this caveat and to avoid introducing fresh terminology,
f (t;  ; ) will be referred to as a characteristicn function
 R tthroughout. Notice that o
we could have started with the
+
joint characteristic function f~(t;  ; ; ')  EtQ exp i' t r(u) du + i  X (t +  ) which is the futures, marked-
 R t+ 
to-market, price of a claim that pays exp i ' t r(u) du + i  X (t +  ) at time t +  . Clearly, the entity in (6)
R
is a special case with f (t;  ; ) = f~(t;  ; ; i). As we will see, all claims contingent on tt+ r(u) du and X (t +  )
are in the span of f~(t;  ; ; '), but are not necessarily spanned by f (t;  ; ). To allow for condensed discussion,
concentrate solely on examining the implications of the characteristic function in (6). When pricing correlation
options, we study this abstraction again.

6
mathematically more tractable: the former being smooth and in nitely di erentiable, while the
latter fails to be di erentiable. As we will establish, the only challenge remaining is to analytically
determine the characteristic function.
For future reference, let Re[:] denote the real part of the expression and L1 (C 2) the space of
integrable (twice continuously di erentiable) functions. Formally, the payo function H(X ) is said
R
to be of class L1 if 11 j H(X ) j dX < 1. Observe that the call payo is L1 modulo an ane
position (i.e., max(0; X K ) (X K ) = max(0; K X )). Motivated by such an implication
de ne, for some constants b and x, universal payo 's of the type
n o
G  H(X ) j H(X ) b x X 2 L1 ; (9)

which encompasses payo functions of wider appeal. With this said, we now compare the span of
eiX and max(0; X K ) and analyze the methodology from di erent perspectives.
Theorem 1 The following relationships hold in arbitrage-free economies:
(a) For generic claim payo 's G , the continuum of characteristic functions (indexed by )
and the continuum of options (indexed by K ) are equivalent classes of spanning secu-
rities. Thus, there exist coecients w() 2 L1nand z (K ) such that: H(X ) = b + x X + o
R 1 Re hw() ei X i d , or H(X ) = lim R1 N
1 N !1 N N
b + x X + 0 z (K ) max(0; X K ) dK
with convergence in the L1 norm; and Nb , Nx and z N (K ) are as displayed in (67)-(69)
of the Appendix.
(b) The call price (in (5)) can be unbundled into a portfolio of Arrow-Debreu securities

C (t;  ; K ) = G(t;  ) 1(t;  ) K B(t;  ) 2(t;  ); (10)

where B (t;  ) is the time-t price of a discount bond with  periods remaining to expira-
tion; G(t;  ) represents the time-t price of a commitment to deliver,
R exp  t +  , the
R t+ r(atu) dutime-
quantity X (t+ ) (scaled forward price); and 1 (t;  )  RX  R tt+  X (t+ ) q() d
exp t r(u) du X (t+ ) q( ) d
R exp R t+ r(u) du q() d
and 2 (t;  )  RX  R tt+  are, respectively, the time-t prices of Arrow-
exp t r(u) du q( ) d
Debreu securities.
(c) Each constituent security required for option valuation in (10) can be recovered from
the characteristic function f (t;  ; ) as follows:

7
  The discount bond price and the scaled forward price respectively obey
B(t;  ) = f (t;  ; 0); (11)
G(t;  ) = 1i  f(t;  ; 0); (12)

where f (t;  ; ) denotes the partial derivative of f (t;  ; ) with respect to ;

 The time t price of each Arrow-Debreu security, for j = 1; 2, is
1 1 Z 1 " e i K  fj (t;  ; ) #
j (t;  ) = 2 +  Re
i d: (13)
0
The characteristic functions for Arrow-Debreu securities, fj (t;  ; ), for j = 1; 2,
are determined from f (t;  ; ) as made exact below

f1(t;  ; ) = i G(1t;  )  f(t;  ; ); (14)

f2(t;  ; ) = B(t;1  )  f (t;  ; ); (15)

where it is understood that f (t;  ; ) is available in closed-form by solving the valuation

equation (8) or the conditional expectation (7).
The upshot that continuum of characteristic functions and continuum of options are equiva-
lent classes of spanning securities in the space of L1 plus ane positions is perhaps not surpris-
ing: the payo on trigonometric functions (options) can be synthesized from options (trigono-
metric
n functions). One can also h envision i thiso portion of the Theorem saying that the residual
R 1
H(X ) b x X 1 Re w() e d is approximately zero in measure{theoretic sense,
i X
and likewise for options (e.g., as in Green and Jarrow (1987); Nachman (1988); and Ross (1976)).
If b = x = 0 and hence H(X ) 2 L1 , from Fourier theory, the static policy in the contin-
R
uum of characteristic functions is complex valued w() = 21 11 H(X ) e i  X dX . Granted that
w() = w1() + i w2(), the admissible trading strategy involves combining (1) a long position
w1() in cos(X ); and (2) a short position w2() in sin(X ) for each . Substituting the Fourier
coecients into the spanning relationship and exploiting the linearity of the martingale pricing
rule, the arbitrage-free value of any claim, in terms of the characteristic function, then becomes
1 R 1 R 1 Re hf (t;  ; ) H(X ) e i X i dX d. Otherwise, if b 6= 0 and x 6= 0, the claim value
2 1 1
should be adjusted by the time t price of the underlying and the discount bond accordingly. Equa-
tion (62) and (66) of the Appendix respectively reveals how H(X ) 2 C 2 and H(X ) 2 L1 can be
synthetically constructed from call options. In spanning the former group of claims, there exists

8
linear combinations in the Lebesgue continuum of strikes; however, in the latter, spanning is in
the L1 norm. In the event calls and characteristic functions are redundant, z (K ) = 0 for all K
and w() = 0 for all .
Central to the methodology, the second part of the Theorem asserts that the call option
price can be decomposed into a portfolio of Arrow-Debreu securities. It implicitly maintains
that knowing the price of four primitive securities (i.e., the matching discount bond, the scaled
forward price, and the two Arrow-Debreu securities) is equivalent to solving the option valuation
problem. To brie y see the logic behind
 R t+this decomposition,
 by the de nition of state-price
R 
density, notice that B (t;  ) = exp t r(u) du q ( ) d and the scaled forward price is:
n  R  o R  R 
G(t;  )  EtQ exp tt+ r(u) du X (t +  ) = exp tt+ r(u) du X (t +  ) q( ) d . By
appealing to the same deduction (and using (3)), one can rigorously represent
R exp  R t+ r(u) du X (t +  ) q( ) d
1 (t;  ) = RX  R tt+  ; (16)
exp t r(u) du X (t +  ) q ( ) d
 EtQ f1X g (17)

where the indicator function 1X is unity when X (t +  ) > K and  Rzero otherwise.
 In deriving
t+
 exp t r(u) du X (t+ )
(17), we have utilized the Radon-Nikodym derivative dQ dQ = G(t; ) . Clearly,
1 (t;  ) is the price of an Arrow-Debreu security, albeit, under a transformed equivalent probability
measure. By an analogous argument
R exp  R t+ r(u) du q( ) d
2(t;  ) = RX  Rtt+  ; (18)
exp t r ( u ) du q (  ) d
Q 
 Et f1X g (19)
 
exp R t+ r(u) du

is a well-posed Arrow-Debreu security with dQdQ = t
B (t; ) . As our reliance is on
elementary properties of probability density functions, the option decomposition holds for arbitrary
risk structures and is valid to valuation problems in discrete-time or in continuous-time{a feature
that can only induce broad theoretical applicability of Theorem 1.4
4 Just consider a claim in a two-period, three date, model written on a non-traded underlying like the price of
electricity at date two. Call this uncertainty x with density q(x) and characteristic function f (x; ). For simplicity,
assume deterministic interest rates. Then, one can verify that the core analysis of Theorem 1 goes through. In non-
Markovian, jump-di usion, or pure-jump environments, the theoretical developments are essentially similar. As our
goal is to avoid repetition, such extended analysis is excluded. See Due, Pan, and Singleton (1998), who provide
a more technical treatment on the determination of 1 (t;  ) and 2 (t;  ) in the context of ane jump-di usions.

9
 The nal part of the Theorem is the real driving force behind the valuation approach, how-
ever. Consistent with the task at hand, the manipulation of the characteristic function f (t;  ; )
simultaneously and jointly recovers (i) the term structure of interest rates; (ii) the term struc-
ture of forward prices; and (iii) the two required Arrow-Debreu securities. While it is known
that options are market completing from Ross (1976), the spanning properties of the character-
istic function are not that transparent and not fully appreciated in their entirety. In fact, as
f (t;  ; )  R eiX (t+ ) exp Rtt+ r(u) du q( ) d , economically it amounts to a Fourier trans-
form of the state-price density function and hence basis augmenting. In particular, the resulting
basis is endowed with two theoretically appealing properties: It is (1) analytical and (2) orthonor-
mal (in L2 ([0; 2 ]) and in the space of almost periodic functions). Observe that by translating or
di erentiating the characteristic function, one can synthesize the values of exponential and poly-
nomial of the underlying uncertainty. The reference measure, for example, is being transformed
from, say, q ( ) to X q ( ) on di erentiation; and to eX q ( ) on translation. Which is precisely the
reason why 1 (2) can be priced by di erentiation (translation) and Fourier transformation (see
also Cases 1 and 2 to follow). These simpli cations are made achievable without deriving the
state-price density function (which is in principle inferable). Moreover, as di erentiation holds
the key to constructing polynomial basis, the superiority of characteristic functions as a primary
collection of spanning securities is evident. After all, the reverse construction contains in nite
series summations.
Theorem 1 should not be interpreted to mean that call options are in the span of trigono-
metric functions via Fourierh theory. Nowhere
i have we established the spanning representation:
R
max(X K; 0) = 11 Re w() ei X d , for some w(). In fact, this is certainly not even
possible using integrable w() as the call payo is unbounded and outside of L1 of Lebesgue
measure. From characteristic functions, one can nonetheless build a large class of functions and
also value them. Because the put option payo is in L1 , it is however algebraically spanned:
there exists linear combinations in the Lebesgue continuum of transform variates . That is
R
w() = 21 11 max(0; K X ) e i X dX . As the call payo is L1 modulo X K , it can therefore
be tailored by investing in (i) the continuum of characteristic functions, (ii) the underlying, and
(iii) the discount bond. The precise long position w1() in cos(X ) and the short position w2() in
sin(X ) that mimic the put payo are displayed in (60)-(61) of the Appendix. By the same token,
reformulating the delta security payo as: 1 1fX<K g does not contradict the impression that
1 and 2 can be synthesized from continuum of characteristic functions in collaboration with
a discount bond (even though each security payo violates L1 requirement). Thus, the formula
(13) is a mere by-product of spanning and pricing via characteristic functions. Such ane payo s
as the discount bond (the underlying) can be specialized from the trigonometric payo by setting

10
 = 0 (di erentiating and then substituting  = 0). In sum total, characteristic functions are
robust spanning engines not only for payo 's in L1 , but also in the expanded collection of L1 plus
ane security positions.5
That the characteristic function and consequently all contingent claims in its basis can be
priced by solving a single valuation equation (partial or integro-di erential) is methodologically
important.6 The reader may recall that the traditional approach to contingent claim/option
valuation pioneered by CIR (1985) and Merton (1973) involves developing a fundamental valuation
equation (such as the one posited in (5)). While circumventing the need to solve for the state-price
density, it, nonetheless, demands a correct candidate conjecture. By analogy, the conjecture is in
the family of (10). Substituting the conjecture into the fundamental equation generally produces
at most four additional valuation equations or the corresponding conditional expectations (i.e., one
each for G(t;  ), B (t;  ), 1 and 2 ). But claim valuation is not yet entirely complete as one must
now again, by trial and error, conjecture a solution to each of the four valuation equations.7 While
the four-step valuation methodology is technically correct and has lead to numerous theoretical
advances and model re nements, it is, nevertheless, cumbersome and imposes a tight constraint
on the valuation structure: if one component valuation is unsolved, it gridlocks contingent claims
5 A few abstract questions have admittedly been bypassed in our inquiry: (1) What is the exact span of character-
istic functions and options? and (2) What is the relationship between the algebraic span of options and characteristic
functions? Under what circumstances is one span larger, or smaller, than its counterpart? For example, X 2 2 C 2
(and e 2 C 2 ) is in the algebraic span of options but not so for characteristic functions using L1 . Yet, if attention is
ex

restricted to some compact interval ( `; `), then, from Fourier theory, X 2 will also be in the span of characteristic
functions as well. Stated di erently, a wider net of securities canRbe spanned by the characteristic function on
compact intervals. In particular, claims that belong to L1 (Q) (i.e., 11 j H(X ) j dq < 1) can be approximated in
the L1 (Q) norm by claims possessing compact support. Thus, contingent claims satisfying this working criteria can
be spanned and priced as depicted above. Evidently, X 2 eiX and successive characteristic function derivatives are
not an L1 object for all , but can be valued anyway from the characteristic function (6).
6 It has been pointed to us that the scaled forward price, G(t; ), is predetermined for a broad class of option
contracts. But as a general rule, this notion is awed. Options on futures (under random volatility and random
interest rates/convenience yields) are an obvious counterexample. Likewise, for the entire family of interest rate
options and nonstandard contingent claims, the scaled forward price is anything but known a priori. To guide
consensus, it is demonstrated in the later example exercises that the scaled forward price embedded in the average
interest{rate and knock-out options with payo , say, Nn=1 max(0; s(t + n
t) K ), are hard to conjecture. While
a large literature exists on the term structure of interest rates, the spirit of the above remarks equally applies to
discount bond prices (although to a lesser extent). Nonetheless, a systematic way to determine the scaled forward
price and the discount bond price is desirable and warranted. In particular, the characteristic function of the
stochastic discount factor can serve a similarly useful role in dynamic equilibrium economies. Details are omitted
here.
7 To clarify this feature of their methodology more formally, remember the solution steps in CIR (and also
Constantinides (1992) and Longsta and Schwartz (1992)). In their well-known model, the discount bond price and
the European option written on it satis es the same fundamental partial di erential equation (hereby, PDE). For
the option formula postulated in CIR to be internally consistent with the valuation PDE (i) the (two) non-central
chi-squared distributed probabilities; and (ii) the (two) discount bond prices will satisfy unique valuation equation
of their own (with a distinct boundary condition). Each component valuation security price was then explicitly
computed by solving the relevant expectation.

11
valuation. In contrast, the availability of the characteristic function renders claims valuation
complete in the same single step (and hence weeds out solving complex valuation equations).
On a related theme, notice that aliated with each constituent security valuation is also a
set of ordinary di erential equations (after a solution is conjectured). Adopting the spanning and
pricing strategy will also eliminate the need to solve a large set of ordinary di erential equations.
Translating and di erentiating a smooth function is clearly trivial relative to solving additional
valuation equations (PDE or integro-di erential) or additional ordinary di erential equations.
But keep in mind that our simpli cations do not apply to the characteristic function of the state-
price density which must be available in closed-form by solving the valuation equation (8) or the
conditional expectation (7). However, due to the characteristic function's exponential boundary
condition, this quantity is easier to solve in general than the option price directly.
While sharing with Heston (1993) the feature that each pure security price is reverse{engineered
from the respective characteristic function, the treatment here departs fundamentally. The two
characteristic functions are, for instance, obtained mostly by solving two separate valuation equa-
tions and by conjecturing their solutions (see equations (12) and (22) in Heston); Whereas, under
our technique, their recovery is through the characteristic function of the state-price density. Our
economic analysis makes explicit how the two characteristic functions are intrinsically linked: the
rst characteristic function is either a translate, or a derivative, of its counterpart. More speci -
cally, existing works tend to blur the recursive structure of option valuation; Seldom has it tapped
into the unifying spanning concept. In this regard, there are crucial lessons to examining claims
with payo s that are variants of the original one in equation (1). Under the Heston framework, the
entire set of valuation equations must be resolved all over again (including a conjecture to the orig-
inal valuation equation). This is, however, not the case under our derivative valuation approach.
Knowing the characteristic function of the state-price density will automatically determine the
intrinsic value of the cash{ ow streams, as is demonstrated below now:
p
CASE 1 Let the claim payo be max(0; X 2(t +  ) K ) with exercise region X (t +  ) > K .
Despite this nonlinear transformation, the claim price still satis es (10) with (as before): B (t;  ) =
f (t;  ; 0) and f2(t;  ; ) = B(1t; ) f (t;  ; ). In accordance with Theorem 1

G(t;  ) = i12 f(t;  ; 0); (20)

f1(t;  ; ) = i2 G1(t;  ) f(t;  ; ); (21)

12
where f (t;  ; ) denotes the second{order partial derivative of f (t;  ; ) with respect to  and

1 1 Z 1 " e i pK  fj (t;  ; ) #
j (t;  ) = 2 +  Re
i d for j = 1; 2:
0
The option claims on successive (higher) algebraic moments and other (integer) polynomials can
be priced correspondingly.
CASE 2 Alter C (t + ; 0; K ) = max(0; eX (t+ ) K ). Here, B(t;  ) = f (t;  ; 0) and f2(t;  ; ) =
1
B (t; ) f (t;  ; ) with

G(t;  ) = f (t;  ; i) ; (22)

f1(t;  ; ) = G(t;1  ) f (t;  ;  i) ; (23)
1 1 Z 1 " e i ln[K]  fj (t;  ; ) #
j (t;  ) = 2 +  Re
i d for j = 1; 2: (24)
0
If we set X (t +  ) = ln[S (t +  )], then (22)-(24) accommodates, as a special parametric case, most
equity option models with f1 (t; ; ) and f2 (t;  ; ) translated (such as the ones in Black-Scholes;
Bakshi, Cao, and Chen (1997); Due, Pan, and Singleton (1998); Heston (1993); Hull and White
(1987); Scott (1997); and Stein and Stein (1991)).
CASE 3 To see the comprehensive nature of the approach, consider an arbitrary option-like payo
max(0; H [X (t +  )] K ) for some (di
 Rerentiable) H [Xi ]X>(t+0.) This contract, however, mandates
R
the knowledge of M (t;  ; )  exp t+ 
t r(u) du e H [X (t+ )] q( ) d . The Appendix
substantiates how M (t;  ; ) and the price of the call is inferable from f (t;  ; ):

G(t;  ) = M (t;  ; 0); (25)

1 1 Z 1 " e i H 1[K]  M (t;  ; ) #
1 (t;  ) = 2 +  Re
i   M (t;  ; 0) d; (26)
0
1 1 Z 1 " e i H 1[K]  f (t;  ; ) #
2 (t;  ) = 2 +  Re
i   f (t;  ; 0) d; (27)
0
where B (t;  ) = f (t;  ; 0) and
X1 A B
n n
M (t;  ; ) = f(t;  ; )  H (X0) + i n  n! (28)
n=1
n
with An  @@Xn Hn (X0); Bn  @ f@(t;n ;) ; and f(t;  ; )  e iX0 f (t;  ; ) is a stand-in for the char-
acteristic function of the translated uncertainty: X (t +  ) X0 (for some constant X0). To x

13
ideas, suppose H [X ] = X (t +  ) for 1 <  < 1. Then, all fractional power claims can be
priced explicitly in closed-form (and other similarly rich arbitrary claims on H [X ] with K = 0).
Regardless of how our alternative approach is interpreted, it brings out a valuation aspect of
immense practical interest. That is, it lls in the gap by making explicit the technical conditions
under which the two probability elements can be members of similar, or distinct, parametric
classes. For a general class of cases, 1 and 2 are connected in a precise way, and this quantitative
relationship is characterized best in terms of the transforms of the state price densities. Speci cally,
this consists of situations when the payo function on which the option is written say, a positive
function, H [X ], is functionally related to a monotone transformation of the underlying uncertainty,
say h[X ]. Hypothesize
X
N X
J
H [X ] = n h[X ]n + j e j h[X ] : (29)
n=1 j =1
R  R t+  i h[X ]
In this case, f (t;  ; )  exp t r(u) du e q( ) d , f2(t;  ; ) = ff((t; ;)
t; ;0) , which allows
us to deduce the general restriction
8N 9
< X n X J =
f1(t;  ; ) = iBG((t;t;)) : inn  @ f2@
(t;  ; ) +
n j f 2 ( t;  ;  i j );; (30)
n=1 j =1

where n , j , and j are arbitrary constants with

8N 9
<X n @ n f2 (t;  ; 0) X J =
G(t;  ) = B(t;  ) : in  @n + j f2 (t;  ; i j ); : (31)
n=1 j =1

In particular, when the option claim is on the exponential of the uncertainty (i.e., h[X ] = ln(X ),
n = 0 for all n, 1 = 1 = 1, and j = 0 for j>1), the characteristic function corresponding
to 1 and 2 are translates of one another; And, which is why 1 and its counterpart inherit
the same parametric class. This property, for instance, induces probability structures in CIR
(which are non-central chi-squared) and stochastic volatility equity option models (the character-
istic functions are each exponential ane) that fall internally in the same family of distributions.
But, when the claim is contingent on the level of the uncertainty (or its powers and polynomi-
als), the rst characteristic function is obtained by di erentiation and the second by translation.
Derivatives so priced can, thus, only be composed of Arrow-Debreu securities that are dissimilar
in their probability-theoretic foundations. To reverse the situation, now keep H [X ] unrestricted
but specialize h[X ] = X . Equations (26){(28) of Case 3 re-illustrate the same dichotomy: the

14
transform of 1 is analytically linked to its counterpart as its consecutive derivative (by replacing
f (t;  ; ) with B(t;  ) f2(t;  ; )).8 For such properties as bounds on delta's and state-prices in
one-dimensional di usion economies, see Grundy and Wiener (1996).
Having said this, we move on to the pricing of speci c contingent claims. Each option-like
security is novel and share a common denominator: no closed-form solutions have yet been discov-
ered (to our knowledge), even though the characteristic function (of the remaining uncertainty) is
straightforward. These derivative securities are all intended to capture the essence and richness
of the spanning induced simpli cation.

2 Average-Rate Interest Rate Options

Inspired by the preceding analysis, the remainder of this section documents how a broad class
of path{dependent claims can be valued using our methodology. To maintain sharp focus, adopt
a payo structure that is average{interest rate contingent. Set the initial date for the averaging
interval to be time{0 (with  no1 loss of generality) and specify the payo on the average-rate call
R
as: C (t + ; 0; K ) = max 0; t+ 0t+ r(u) du K where the time{t call option price is denoted
by C (t;  ; K ). To avoid free-lunches
  Z t+   1 Z t+ A ( t )  
C (t;  ; K ) = EtQ exp r(u) du  t +  r(u) du + t +  K  1E ; (32)
t t
where 1E stands for an indicator variable which is unity when the call is exercised (and zero
R R
otherwise); E  f tt+ r(u) du > [t +  ] K A(t)g; and A(t)  0t r(u) du. Therefore

dA(t) = r(t) dt (33)

from Leibnitz's di erentiation rule.

For its theoretical tractability, assume that the law of motion for the spot interest rate, r, is
governed by the single-factor CIR{type square{root process (, , and  are all positive constants):
q
dr(t) =  ( r(t)) dt +  r(t) d!r (t): (34)

Because (r(t); A(t)) form a Markov system from (33)-(34), standard steps produce the valuation
8 At a conceptual level, Cases 1 and 3 highlight a subtle, yet crucial, attribute of the valuation paradigm: the
delta claim 2 is comparatively easier to determine than 1 . In other words, when 1 and 2 lie in di erent
parametric classes, it is trickier to guess solutions to f1 (t; ) and G(t;  ) and hence to the composite option problem.
Consequently, when pricing non-traditional and exotic derivatives, our simpli cation is more about determining 1
and G(t;  ) rather than 2 . We revisit this theme when pricing average-rate interest rate claims.

15
PDE for the average-rate call displayed below (subject to the call payo ):9
1  2 r C +  ( r) C C r C + r C = 0; (35)
2 rr r  A

where the subscripts Cr and Crr respectively represent, for instance, the rst and the second{order
partial derivative with respect to r. Two points are worth mentioning. First, as the evolution of
r(t) is under the martingale measure, the interest rate factor risk premium is already re ected
in the drift:  ( r). With A(t) deterministic, no risk compensation is required for this state
variable. Second, by virtue of its dependence on A(t), the option price is path-dependent. As
a consequence, the valuation equation (35) di ers from its now famous counterparts (e.g., CIR,
Constantinides (1992) and Longsta and Schwartz (1992), among others). Essentially, the induced
path-dependence has made valuation intractable and no (complete) analytical characterizations
have yet been proposed (for single or multi-factor interest rate processes).10
Despite the hurdles in solving (35), directed by Theorem 1, the characteristic function is the
sole building block for spanning and pricing all average-rate contingent claims. To articulate this
point in sucient detail, it is rst veri ed in the Appendix that
  Z t+   Z t+ 
f (t;  ; )  EtQ exp r(u) du  exp i  r(u) du
t t
= exp [ M( ; ) N ( ; ) r(t)] ; (36)

where M( ; ) and N ( ; ) are de ned in (85) and (86) of the Appendix. Next, relying on a
parallel theoretical development and (36), the solution to (35) is as (recursively) posited below:
Proposition 1 The call option price on the average-interest takes the form:
G(t;  )  A(t) 
C (t;  ; K ) = t +  1(t;  ) K t +  B(t;  ) 2(t;  ); (37)

where B (t;  ) = f (t;  ; 0) = exp [ M( ; 0) N ( ; 0) r(t)] with f (t;  ; ) as postulated in (36); and
9 Ifany alternative single{factor or multiple{factor interest rate model is used as a benchmark instead, the
characteristic function will be slightly more dicult to solve analytically (e.g., Constantinides (1992); Longsta
(1989); or in the Markovian jump-di usion mixture class). But as can be seen, the main thrust of this section is
invariant to any one such choice.
10 See the recent contributions (partial list) by Bakshi and Madan (1997), Chacko and Das (1997), Geman and
Yor (1993), Ju (1997) and Zhang (1998) on how average-rate derivatives are routinely adopted by practitioners to
manage interest rate and commodity price risk. In relation to existing research, the incremental contribution of this
paper will be noted shortly.

16
the time{t scaled forward price is

G(t;  ) = 1i  f(t;  ; 0)
1  @M @ N r(t)
= lim
!0 i
 exp [ M( ; ) N ( ; ) r(t)]  @ @ (38)

M and @ N ; and the time{t price of delta securities, for

for some (easily computable) functions: @@ @
j = 1; 2, are Z 1 " e i [(t+ )K A(t)]  fj (t;  ; ) #
1 1
j (t;  ) = 2 +  Re d; (39)
0 i
with the rst characteristic function determined from

f1(t;  ; ) = i G(1t;  )  f(t;  ; )

1  @M @ N r(t) ;
=
i G(t;  )  exp [ M( ; ) N ( ; ) r(t)]  @ @ (40)

and the second characteristic function is

f2(t;  ; ) = B(t;1  ) exp [ M( ; ) N ( ; ) r(t)] : (41)

The formula (37) constitutes an exact closed{form solution to the option on average-interest.11
It brings into forefront the valuation role of the characteristic function; By its translation and
di erentiation, all the underlying primitive securities can be recovered. To see how the spanning
and pricing engine works in practice and to assess the extent of simpli cation it induces, take the
rst Arrow-Debreu security. Write the valuation PDE/Backward-equation (see (81)) as
1  2 r @ 21 +  ( r) +  2 r 1 @G  @ 1 @ 1 + r @ 1 = 0: (42)
2 @r2 G @r @r @ @A
At rst glance, it appears that no closed-form representation may be possible for this class of
PDE's. But as made precise in (40), by di erentiating the characteristic function and standardizing
the resulting entity by G(t;  ) (which makes it a characteristic function for a probability) and
using the inverse Fourier{transformation pins-down the solution to the valuation equation (42).
11 Under the premise that r is a Bessel process and r(u) are independent for all u, the average-rate call can
be priced via Geman-Yor (1993) (since Bessel processes are stable under additivity). Unfortunately, none of the
existing processes fall into the viability set. When r is governed according to (34),
P it is obviously Bessel; But it is
auto-correlated, however. Now suppose the option is on a basket of securities: Jj=1 j xj (t +  ) (for some loading
j ); and xj (t)'s are all independent Bessel, then it poses no valuation diculties (via our approach or Geman-Yor).

17
This step can be consistently implemented so long as the characteristic function is analytical.
But the two modes of analysis are not strictly equivalent: one requiring a simple di erentiation
step; the other, intricate conjecturing abilities. Proceeding similarly, the scaled forward price is in
compliance with: 21  2 r Grr +  ( r) Gr G r G = r B (t;  ); and its closed-form formulation
in (38) stems virtually from the same mechanism. In words, by di erentiating the characteristic
function and evaluating the resulting expression at zero replicates the conditional expectation for
this vanishing contingent claim (notice that the price is monotonically declining n Awith o the passage
of time). The pricing of the put can be achieved from put{call parity: K t+ B (t;  ) [1 (t )
2 (t;  )] Gt(+t; ) [1 1 (t;  )].12
Proposition 1 imposes a stringent restriction on the pure securities: 1 and 2 . This is
primarily so since f2 (t;  ; ) is exponential{ane, but f1 (t;  ; ) is surely outside of that class.
Furthermore, 2 (t;  ) is non-central chi-squared distributed (it satis es the Kolmogorov-Backward
equation for non-central chi-squared variable, as in CIR bond option formula). Thus, the average-
rate option valuation problem is potentially one application in which 1 and 2 are not in the
same parametric family of distribution functions.13
The principal comparative statics ndings are not at odds with what intuition suggests: the
average-rate call is (1) increasing in A(t) (the prior path-dependence); and (2) convex and increas-
ing in r(t). Numerical analysis indicates that higher interest rate primitives (i.e., , , and  ) all
lead to a higher call price. By shorting N (t;  ; 0) Cr units of the discount bond and going long
C (t;  ) + N (t;  ; 0) Cr B(t;  ) in cash (which makes the overall position self{ nanced), the call can
be dynamically replicated. Because the partial derivative Cr is in analytical{form, the closed-form
characterization facilitates the execution of delta{neutral hedges.
Our approach can be adapted to price assorted payo structures. To synthesize this dimension
of the technique, take the option on the average-yield as the basis. Maintaining the CIR interest
rate dynamics, the yield-to-maturity R(t; ~) = M(~~ ;0) + N (~~ ;0) r(t). Let C (t;  ) denote the call
option price on this average with expiration  {periods from time t. Then

C (t;  ; K ) = (Nt +(~; 0)

)~ G ( t;  ) 1 ( t;  ) K B(t;  )  (t;  );
(t +  )~ 2 (43)

12 By and large, the methodology o ered here to price average-rate claims is superior and mathematically more
elegant
R than in Ju (1997) (because of spanning). In his innovation, Ju analytically derives the Fourier transform
of tt+ r(u) du. Then, via inverse transformation, he recovers the state-price density function (see equation (24) in
his paper). So, his solution for average-call is numerical in nature. Nonetheless, the primary message is that the
state-price density is generally redundant for derivative asset valuation, if the characteristic function is known.
13 In attacking the exact same problem, Chacko and Das (1997) were unable to analytically characterize G(t;  )
and f1 (t;  ), which is a consequence of our approach. As a result, their general focus is con ned to the pricing of
the Digital (i.e., 2 (t;  )). Thus, on balance, when one conjectures ad-hoc solutions to option valuation problems,
it tends to obscure the tight linkage between each Arrow-Debreu security and between G(t;  ) and f1 (t;  ).

18
where K  ~ [t +  ]K [t +  ] M(~ ; 0) N (~ ; 0) A(t); and the risk{neutralized probabilities are
1 1 Z1 1 iK

j (t;  ) = 2 +  Re i e N (~ ;0) fj (t;  ; ) d j = 1; 2; (44)
0
where G(t;  ), f1 (t;  ; ), f2 (t;  ; ) are as respectively given in (38) and (40){(41). Option claims
on (i) higher algebraic moments and (ii) fractional powers and polynomials can also be accom-
modated in the same single-step. In summary, by constructing the appropriate characteristic
function, any interest rate derivative can be priced in closed-form.

3 Correlation Options
The preceding contingent claims were all written on a single underlying asset (and with a one-
dimensional exercise region). However, valuation applications of special interest to nancial
economist often have payo 's dependent on two assets. To see how such claims can be priced
within our framework, let s(t) and p(t) be the time t price of the two securities. Specify a generic
payo of the type:
   
C (t + ; 0; Ks; Kp) = max 0; s(ts(+t) ) Ks  max 0; p(pt (+t) ) Kp (45)

where C (t + ; 0; Ks; Kp) is the price of the correlation option at time{t +  and the respective
strike prices are denoted by Ks and Kp . Since 1 s(t+ ) >Ks  1 p(t+ ) >Kp = max(0; 1 s(t+ ) >Ks +
s(t) p(t) s(t)
1 p(t+ ) >Kp 1), it is noteworthy, from Nachman (1988) and Ross (1976), that correlation options
p(t)
are market completing. The key result of this section is stated next.
Proposition 2 Let s(t) and p(t) each be governed by the continuous-time stochastic processes
(under the risk-neutral measure) below
ds(t) = r dt +  qv(t) d! (t);
s(t) s s
q
dp(t) = r dt +  v(t) d! (t);
p(t) p p
q
d v(t) = [ v(t) ] dt +  v(t) d!v (t);
where Covt (!s (t); !p(t))   , Covt (!s (t); !v (t))  1, Covt (!p (t); !v (t))  2, and r is the con-
stant interest rate. For this problem, the joint characteristic function is:
   s(t +  )   p(t +  ) 
f (t;  ; ; ') = EtQ exp r + i  ln s(t) + i ' ln
p(t) (46)

19
 = exp [ Y (t;  ; ; ') + Z (t;  ; ; ') v(t) ] ; (47)

where Y (t;  ; ; ') and Z (t;  ; ; ') are displayed in (88)-(89) of the Appendix. Then

C (t;  ) = f (t;  ; i; i) 1(t;  ) Ks 2(t;  ) Kp 3(t;  ) + KsKp e r 4 (t;  ); (48)

 h i h i 
where j (t;  )  Prob fln s(st(+t) ) > ln[Ks]g \ fln p(pt(+t) ) > ln[Kp]g (under mutually equivalent
probability measures) with, for j = 1;


; 4:
"
Z 1 e i ln[Ks] fj (t;  ; ; 0) # "
Z 1 e i' ln[Kp] fj (t;  ; 0; ') #
1 1
j (t;  ) = 4 + 2 Re
1
d + 2 Re d'
0 i 0 i'
( " # " #)
1 Z 1 Z 1 Re e i ln[Ks] i' ln[Kp ] fj (t;  ; ; ') Re e i ln[Ks]+i' ln[Kp ] fj (t;  ; ; ') dd':
2 2 0 0 ' '
The corresponding characteristic function are:

f1(t;  ; ; ') = f (t;  ;  i; ' i) ;

f (t;  ; i; i)
f2(t;  ; ; ') = f (t;  ; ; ' i) ;
f3(t;  ; ; ') = f (t;  ;  i; ');
f4(t;  ; ; ') = er f (t;  ; ; ') :
 s(t+ ) max 0; K p(t+ ) 
The price of the put option with payo : C (t+; 0; Ks; Kp) = max 0; Ks s(t) p p(t)
can be deduced from put-call parity for correlation options.
In extending existing treatments, our work provides at least three additional contributions.
First, we o er a exact closed{form solution for correlation options under stochastic volatility. By
setting = =  = 0 and using L'Hopital's rule, the valuation formula in (48) converges to the
counterpart one under geometric Brownian motion. Under these restrictions, the return charac-
teristic functions in (46){(47) are precisely those of the bivariate normal distribution. Having a
more general valuation formula where (i) shocks to each asset are correlated with volatility shocks
(i.e., 1 6= 0 and 2 6= 0) and (ii) the modeling of a more plausible and time-varying correlation
structure between assets should help close the gap in understanding and predicting how these
claims respond in a non-lognormal setting. The general stochastic structure considered in Propo-
sition 2 is consequently more consistent with such real{life applications as contingent securities on
currency bonds, commodity{linked bonds and on cross-exchange rates (and where market forces
induce stochastically-varying asset price co-movements).

20
 Second, by the spanning property of the joint characteristic function, valuation again has a
one-step avor. That is, the closed{form expression for f (t;  ; ; ') guarantees the simultaneous
recovery of all the (four) characteristic functions and one scaled forward price; As a consequence,
the requirement that the counterpart valuation equations be explicitly evaluated has been by-
passed even under this two-dimensional exercise region set-up (see Stulz (1982) for the latter
approach). Observe that the assumption of time-invariant interest rates and zero convenience
yields is for the sake of expositional convenience only. In such general settings and for other two-
dimensional contract structures, the traditional approach will be far more demanding (at most
eight valuation equations in total). So, the simpli cations via the integrative spanning concept
are likely to be substantive there as well. For instance, option valuation in the maximum or
the minimum of two asset class (e.g., Stulz (1982)) also hinges on the joint characteristic func-
tion (46). As a result, their valuations can be reconciled internally within Proposition 2. In
the same s(t+ ) rising (decling) and p(t+ ) declining (rising):
 s(tspirit,
+  )
payo
 variants
 of (45) with
p ( t + )  s(t) 
s ( t +  )  p(t) p(t+ ) 
max 0; s(t) Ks  max 0; Kp p(t) (or, max 0; Ks s(t)  max 0; p(t) Kp ) in-
volves option{pricing under negative cross-correlation, which is also within the scope of the present
model.14 In fact, the pricing of any claim on the joint uncertainty is immediate from (46)-(47).
Lastly, in deriving the Arrow-Debreu security prices j (t;  ), for j = 1;


; 4, we have intro-
duced the inversion formula for their determination (when the exercise region is a bivariate vector).
Recall that the probability 4 is the time t price of the delta security (under the risk-neutral mea-
sure) that pays one dollar when: ln[s(t +  )] ln[s(t)] > ln[Ks] and ln[p(t +  )] ln[p(t)] > ln[Kp],
and zero otherwise. Other probabilities have similar intuitive interpretations. In proposing our
solutions for the delta securities, we have adapted a result originally due to Shephard (1991) (de-
tails are in the Appendix) which has allowed us to extend the one-dimensional Fourier inversion
methodology (as in Heston (1993); Kendall and Stuart (1977); and Lukacs (1960)) to the pricing
of options written on two assets. Moreover, because s(t) and p(t) are proportional stochastic
processes and the option payo is exponential-ane in the joint uncertainty, all the characteris-
tic functions are translated from the joint characteristic function. Therefore, unlike the previous
example, the probabilities, j , are in the same parametric class. It remains to be emphasized
that although the determination of each probability demands a bivariate numerical integration,
it presents no implementation diculties. In reality, the probabilities and hence the option prices
can be obtained with high speed.
14 Some authors such as Zhang (1998) have described the correlation (call) option contract to mean the following
 
payo : max 0; s(st(+t) ) Ks if p(pt(+t) ) > Kp and vice-versa. Restricting the middle two terms in (48) to zero, and
setting Kp = 1 in the fourth one will give the price of such a contract.

21
4 A Class of Discretely-Monitored Knock-out Options
For this nal application, consider a discretely-monitored knock-out call option with contractually
determined payo (for ease of exposition, some of the notation has changed)

C (t + N
t; 0) = 1fs(t+
t)>Kg  1fs(t+2
t)>Kg 


 1fs(t+N
t)>Kg  max (0; s(t + N
t) K )

(49)
where s(t) is the time-t price of the spot asset. If at any time prior to expiration, the spot price
goes below a pre-speci ed barrier K , the call option is knocked-out and hence worthless. For
simplicity, assume that the spot price evolves according to a log-normal process
p
ln[s(t +
t)] ln[s(t)] = (r 21  2)
t + 
t (t +
t) s(0) > 0; (50)

where r and  are constants and (t) is a standard normal variate for all t. Thus, the characteristic
function of the remaining uncertainty with density q (s(t +
t);


; s(t + N
t)) is
Z1 Z1 XN !
f (1;


; N ) =


e r N
t exp i n ln[S (t + n
t)] dq(s(t +
t);


; s(t + N
t))
0 0 n =1
2 0 T 12 N 3
  XX
N N X X
N X
= exp 64 rN
t + r 21  2
t in + 12 2
t @ itA + in ln[s(t)]75 :
j =1 n=j j =1 n=j n=1

Using a similar sequence of steps as in Theorem 1, we have the call price

C (t; N
t; K ) = f (0;


; 0; i) 1(t; N
t) K e rN
t 2 (t; N
t) (51)

where the characteristic functions corresponding to the risk-neutral probabilities are respectively15

f1(1;


; N ) = f (1;f
(0

;;

N
; 01;; iN) i) (52)

f2(1;


; N ) = erN
tf (1;


; N ) (53)

and the N-dimensional Fourier inversion formula in Shephard (1991) can be adapted to arrive at the
probabilities: 1 and 2. This N-day, or N-week, formula is recursive and easily programmable in
15 One can complement the log-normal assumption by a stochastic process with (i) poisson jump arrival rates and
lognormal/gamma distributed jump intensities, or (ii) pure-jump processes under generalized Levy measures (say,
with in nite arrival rates). Although each enhancement will lead to distributions that dominate the log-normal
pricing distribution on several fronts, their modeling aspects can be substantially involved (especially with the Levy
measure). As this can ultimately be very distracting to the reader, these extensions were largely ignored in the
development of (51).

22
standard packages. For the twin-contract: Nn=1 max(0; s(t + n
t) K ), the structure of valuation
is only slightly more complex (but uses the same joint-characteristic function). Since formulas for
the probabilities are not compact, they are omitted to save on space and to stress the spanning
focal point.

5 Conclusions
While it is known that the value of the call option recovers the corresponding put option price
without actually solving its valuation equation; It is, however, not yet fully understood that the
valuation equation for the characteristic function is sucient to recover all underlying primitive
claim prices. As the main idea is spanning, this remark is valid whether the underlying is an
option on the underlying uncertainty, on its powers, or its exponential, or virtually any other
monotonic function of the uncertainty. Our work also provides a way to formalize and unify the
valuation of the term structure of interest rates, the valuation of the term structure of forward
and futures prices, and the valuation of Arrow{Debreu securities. When derivative securities with
higher{dimensional exercise regions were considered, the characteristic function basis provided
superior analytical tractability (as in correlation and discretely-monitored knock-out options).
Our work can be extended along several dimensions. First, it may used to examine contingent
claims valuation in the context of Heath, Jarrow, and Morton (1992). Adopting characteristic
function-based methods may alleviate the burden of derivative-security valuation in their models.
Second, it may be utilized to revisit American option valuation under more plausible stochastic
dynamics. Here, one may determine the characteristic function of the optimal stopping problem
which may jointly recover the European option price and the early exercise premium. Finally, the
methodology can be used to price exotic options with complex boundary conditions and under
stochastic volatility (e.g., barriers and lookbacks). All of these characterizations are left to a future
scrutiny.

23
 Appendix
Proof of Spanning Equivalence in Part (a) of Theorem 1.
Recall ei  X = cos( X ) + i sin( X ). Thus, the proof entails comparing the span of trigono-
metric functions with those of call options. The proof is divided into four parts for clarity.
(i) Spanning Claims in L1 Via Characteristic Functions
For now, x b = x = 0 and let H(X ) 2 L1 be the claim payo under consideration.
Suppressing time arguments on X , de ne trading strategies as complex valued policies w() 2 L1
such that Z1 h i
H(X ) = Re w() ei X d (54)
1
which is just the cash ow attained by the strategy with
Z1
w() = w1() + i w2() = 21 H(X ) e i X dX (55)
1
by the mathematics of inverse Fourier transformation. Thus, if the portfolio policy implicit in
(54){(55) is adopted, then Fourier theory for L1 asserts that (54) holds exactly (i.e., Goldberg
(1965, Chapter 1)). So, H(X ) is in the algebraic span of trigonometric functions.
Disentangling w() into its real and imaginary components as in (55) and substituting into
equation (54) produces
Z1
H(X ) = Re [(w1() + i w2())fcos( X ) + i sin( X )g] d (56)
Z 11
= [w1() cos( X ) w2() sin( X )] d (57)
1
which formalizes how the continuum of long positions: w1() in cos( X ), and short positions:
w2() in sin( X ) can conceive any H(X ) 2 L1. Simplifying (55)
Z1 Z1
w1() + i w2() = 21 H(X ) cos(X ) dX i 21 H(X ) sin(X ) dX (58)
1 1
determines w1() and w2() in terms of the payo function to be spanned and the sine and the
cosine. This completes the description of how to span claims in L1 via trigonometric functions.
(ii) Spanning Call Options Via Characteristic Functions
The payo on the call option is not in L1 . Exploiting the identity max(0; X K ) = max(0; K
X ) + X K , and Fourier theory applied to the put payo (since max(0; K X ) 2 L1) yields
Z1 h i
max(0; X K ) = b + x X + i X
Re w() e d (59)
1

24
 R
where w() = w1() + i w2() = 21 11 max(0; K X ) e i X dX , b = K , and x = 1. The
exact composition of w1() and w2() remains to be shown. From standard integration steps
ZK
w1() = 21 (K X ) cos(X ) dX
0
= 1 [1 cos(K )] (60)
22
and again from equation (58)

w2() = 1 Z K (K X ) sin(X ) dX
2 0
= 1 K sin(K )  (61)
2  2
which are the legitimate long and short positions in the cosine and the sine to span calls from
characteristic functions (augmented by the discount bond and the underlying).
(iii) Spanning Characteristic Functions Via Call Options
From Theorem 1 of Carr and Madan (1997), the following spanning representation holds for
H(X ) 2 C 2 (the space of twice continuously di erentiable functions):
Z X0 Z1
H(X ) = H(X0)+HX (X0)(X X0)+ HXX (K ) max(0; K X ) dK+ HXX (K ) max(0; X K ) dK
0 X0
(62)
for some constant X0 and HX (HXX ) stands for the rst (second) order partial derivative of the
claim payo with respect to X . Or, by substituting X0 = 0, and H(X ) = cos(X ) delivers
Z1
cos(X ) = b + z(K ) max(0; X K ) dK (63)
0
for b = 1 and z (K ) = 2 cos(K ). Similarly letting H(X ) = sin(X ) and X0 = 0, we recover
Z1
sin(X ) = x X + z(K ) max(0; X K ) dK (64)
0
for x = 1 and z (K ) = 2 sin(K ). Thus, we have the result that trigonometric functions, and
hence characteristic functions, can be e ectively synthesized from a continuum of call options.
(iv) Spanning Claims in L1 Via Call Options
For the purpose of spanning claims H(X ) 2 L1 through call options, we adopt the result due
to Wiener (as exposited in Goldberg (1965, page 32-33)) that H(X ) can be constructed from a

25
 nite portfolio of H (X ) 2 C 2 . Hence in particular:
Z1 X
JN
lim
N !1
H(X ) aNj HN (X + bNj ) dX ! 0 (65)
1 j =1

for complex sequence aNj and real bNj for j = 1;


; J N . To accomplish (65), observe that the
standard Gaussian density has (i) a well-de ned Fourier transform; (ii) at least C 2; and (iii)
satis es all the technical regularity conditions in Theorem 10C of Goldberg. Thus, by taking
2
H (X ) = p12  e X2 , one can span any integrable function by a linear combination of translated
standard Gaussian densities. Implementing these steps in the present context and using (62) to
span H (X ), we can conclude
 Z 
H(X ) = Nlim  N + N X + 1 z N (K ) max(0; X K ) dK (66)
!1 b x 0
in the L1 norm with
X
J N 2
 p12  aNj e
bN
Nb
j
2 (67)
j =1
X
J N 2
p1
bN
Nx  aN bN e
j
2 (68)
2  j =1 j j
X
J N 2
 p12  aNj e
(bN
j +K )
z N (K ) 2 [(bNj + K )2 1] (69)
j =1

and the theorem is proved.

Thus, in summary, we have proved that (i) continuum of options and continuum of trigono-
metric functions are equivalent classes of spanning securities for L1 with convergence in the L1
norm; and (ii) the construction of cash ows has a similar integral representation with the resulting
classes of functions not mutually orthogonal. 2
Proof of Part (b) and Part (c) of Theorem 1.
The proof relies on the fundamental properties of probability density functions and character-
istic functions. By collapsing the integral in (3), it follows that

C (t;  ; K ) = G(t;  ) 1(t;  ) K B(t;  ) 2(t;  ); (70)

26
where the scaled forward price and the discount bond price are
Z  Z t+ 
G(t;  )  exp r(u) du X (t +  ) q( ) d; (71)
Z  Zt t+ 
B(t;  )  exp r(u) du q( ) d; (72)
t
R 
and recalling that q ( ) denotes the risk-neutral density of   tt+ r(u) du; X (t +  ) ; and X
and respectively stand for the sets: fX (t +  ) > K g and fX (t +  ) > 0g. Continuing with the
same style of reasoning
R exp  R t+ r(u) du X (t +  ) q( ) d
1 (t;  )  RX  t
R t+ r(u) du X (t +  ) q( ) d ; (73)
exp t
R exp  R t+ r(u) du q( ) d
2 (t;  )  RX  t
R t+ r(u) du q( ) d ; (74)
exp t

which are valid probabilities since j 2 (0; 1), for j = 1; 2.

To substantiate the relationship between each constituent security in equation (5) and the
characteristic function, f (t;  ; ), de ne
Z  Z t+ 
f (t;  ; )  exp r(u) du ei  X (t+ ) q( ) d; (75)
t
which is the Fourier transform of the state-price density. Setting  = 0 and rearranging
Z  Z t+ 
f (t;  ; 0) = exp r(u) du q( ) d
t
 B(t;  )
which coincides with (11) of Theorem 1. Correspondingly, di erentiating both sides of equation
(75) with respect to 
Z  Z t+ 
f (t;  ; ) = i exp r(u) du  X (t +  )  ei X (t+ ) q( )d: (76)
t
Using (76) and evaluating f (t;  ; ) at  = 0 justi es the assertion in (12).
Now generate the characteristic function of the rst Arrow-Debreu function, f1 (t;  ), from

27
equation (73), as (e.g., Lukacs (1960); and Kendall and Stuart (1977)):
R exp  R t+ r(u) du X (t +  ) ei X (t+ ) q( ) d
t
f1(t;  ; ) = R exp  R t+ r(u) du X (t +  ) q( ) d ; (77)
t
 ff((t;t;  ;; 0)) (78)


with the aid of (76). By similarly manipulating the primitive characteristic function, the derivation
of f2 (t;  ; ) is quite clear-cut. 2
Proof of (25)-(28) in Case 3.
Let X (t +  ) X0 denote translated uncertainty for some constant X0 and f(t;  ; ) its char-
acteristic function. Then f(t;  ; ) = ei  X0 f (t;  ; ). Now
Z  Z t+ 
M (t;  ; )  exp r(u) du H (X ) ei X q( ) d: (79)
t
Taking a Taylor series of H (X ) around X0 and reformulating each component security as in
Theorem 1 con rms (25)-(28). 2
Proof of the Path-Dependent Interest Rate Option Formula in Proposition 1.
R
Proceeding in the same spirit as Theorem 1, now let   tt+ r(u) du with density q ( ) and
E  f > [t +  ] K A(t)g. By de nition
Z   A(t) K  q( ) d:
C (t;  ) = e + (80)
E t+ t+
R
Decomposing this conditionalR expectation, we get (37) with GR (t;  )  01 e   q ( ) d ; B (t;  ) 
R 1 e  q( ) d ;  (t;  )  R 1E e   q() d ; and  (t;  )  R 1E e  q() d . So the corresponding
0 1 e   q( ) d 2 e  q( ) d
0 0
characteristic functions must be:
1 Z1
f1(t;  ; ) = G(t;  ) e   ei q( ) d; (81)
1 Z01
f2(t;  ; ) = B(t;  ) e  ei q( ) d; (82)
0
and the characteristic function of the state-price density is
Z1
f (t;  ; ) = e  ei q( ) d; (83)
0   Z t+   Z t+ 
= EtQ exp r(u) du  exp i  r(u) du : (84)
t t

28
Di erentiating and translating (83) con rms (40)-(41) of Proposition 1. Finally, solving (83) lls
p
in the missing structural link displayed in (36) with ()  2 2(i 1)  2 and
"  ) !#
M( ; )  2 ( ) + 2 ln 1 ( )(1 e
2 ; (85)

N ( ; )  2 2(1 ( i)(1) (1 e e ) ) : (86)

2
Proof of Proposition 2. h i h i
For parsimony of presentation, let S  ln s(st(+t) ) and P  ln p(pt(+t) ) . Omitting time-
arguments, write the joint density function as q (S; P ). The joint characteristic function is
Z1Z1
f (t;  ; ; ')  e r +i S+i ' P q(S; P ) dS dP (87)
1 1
Directly solving this conditional expectation delivers (47) with
" ! #
Y (t;  ; ; ')  [i + i' 1] r  [# ](1 e # ) [# ]
2 2 ln 1 2# (88)
e # )
Z (t;  ; ; ')  2# 2[# (1 ](1 (89)
e # )
p2
and de ning (; ')  i  s  1 i ' p  2; #(; ')  2  2  ; and  (; ')  12 i  s2
1 i ' p2 1 2 s2 1 '2 p2  ' s p  .
2 2 2
Returning to the correlation option problem, rewrite the conditional expectation (45) as
Z1 Z1 n o
C (t;  ; Ks; Kp) = e r +S+P Kse r +P Kpe r +S + KsKpe r q(S; P ) dS dP
ln[Kp ] ln[Ks]
= f (t;  ; i; i) 1(t;  ) Ks 2 (t;  ) Kp 3 (t;  ) + Ks Kp e r 4 (t;  )

and the claim in (48) follows by observing that

Z1Z1
f (t;  ; i; i) = e r +S+P q(S; P ) dS dP (90)
R 11 R11 e r +S+P q(S; P ) dS dP
1(t;  ) = ln[R K1p] Rln[ Ks]
1 e r +S +P q (S; P ) dS dP (91)
1 1
R R R R
and 11 11 e r +S q (S; P ) dS dP = 11 11 e r +P q (S; P ) dS dP = 1 (the martingale restric-
tion).

29
 The sole task remaining is to get each probability. Again, focus on the probability 1(t;  ).
By appealing to standard probability theory, one can verify that its characteristic function is:
1 Z1Z1
f1(t;  ; ; ') = f (t;  ; i; i) e r +S+P  ei S+i ' P q(S; P ) dS dP
1 1
 f (t;  ; i; i) i)
f ( t;  ;  i; '

which arms that f1 (t;  ) can be expressed in terms of the joint characteristic function.
Adapting Theorem 5 in Shephard (1991) to the present two-dimensional problem, for any
distribution function F (S; P ; a; b) with joint characteristic function f (S; P ; ; '), the following
can be asserted:

F (S; P ; a; b) = 41 + 12 F (S ; a) + 12 F (P ; b)
" # " #!
1 Z 1 Z 1 Re e ia i'b f (S; P ; ; ') Re e ia+i'b f (S; P ; ; ') d d'; (92)
2 2 0 0 ' '
and the marginal distributions for S and P are, respectively, given by:
" #
F (S ; a) = 12 1 Z 1 Re e ia f (S ; ; 0) d; (93)
 0 i
" #
F (P ; b) = 12 1 Z 1 Re e i'b f (P ; 0; ') d': (94)
 0 i'
Armed by the joint and marginal distributions in (92)-(94), the probability 1 (t;  ) can now be
constructed as follows:

1(t;  ) = 1 F (S ; ln[Ks ]) F (P ; ln[Kp]) + F (S; P ; ln[Ks]; ln[Kp])

Rearranging veri es the Fourier-inversion formula displayed in Proposition 2. 2

30
 References

Bakshi, Gurdip, Charles Cao, and Zhiwu Chen, 1997, \Empirical Performance of Alternative
Option Pricing Models," Journal of Finance 52, 2003-2049.
Bakshi, Gurdip, and Dilip Madan, 1997,\Average-Rate Contingent Claims," mimeo, University of
Maryland.
Breeden, Douglas, and Robert Litzenberger, 1978, \Prices of State-Contingent Claims Implicit in
Option Prices," Journal of Business 51, 621-651.
Carr, Peter, and Dilip Madan, 1997, \Optimal Positioning in Derivative Securities," mimeo, Mor-
gan Stanley, and the University of Maryland.
Chacko, George, and Sanjiv Das, 1997, \Average Interest," NBER # 6045.
Constantinides, George, 1992, \A Theory of the Nominal Term Structure of Interest Rates,"
Review of Financial Studies 5, 531-552.
Cox, John, and Stephen Ross, 1976, \The Valuation of Options for Alternative Stochastic Pro-
cesses," Journal of Financial Economics 3, 145-166.
Cox, John, Jonathan Ingersoll, and Stephen Ross, 1985, \A Theory of the Term Structure of
Interest Rates," Econometrica 53, 385-408.
Due, Darrell, Jun Pan, and Kenneth Singleton, 1998, \Transform Analysis and Option Pricing
for Ane Jump-Di usions," mimeo, Stanford University.
Geman, Helyette, and Marc Yor, 1993, \Bessel Processes, Asian Options and Perpetuities," Math-
ematical Finance 3, 349{375.
Goldberg, Richard, 1965, Fourier Transforms, Cambridge University Press, London.
Green, Richard, and Robert Jarrow, 1987, \Spanning and Completeness with Contingent Claims,"
Journal of Economic Theory 41, 202-210.
Grundy, Bruce, and Zvi Wiener, 1996, \The Analysis of VAR, Delta and State Prices: A New
Approach," mimeo, University of Pennsylvania.
Heath, David, Robert Jarrow, and Andrew Morton, 1992, \Bond Pricing and the Term Structure
of Interest Rates: A New Methodology for Contingent Claims Valuation," Econometrica 60, 77-
105.
Heston, Steven, 1993, \A Closed-Form Solution for Options with Stochastic Volatility with Ap-
plications to Bond and Currency Options," Review of Financial Studies 6, 327-343.

31
Hull, John, and Alan White, 1987, \The Pricing of Options on Assets with Stochastic Volatilities,"
Journal of Finance 42, 281{300.
Ju, Nengjiu, 1997, \Fourier Transformation, Martingale, and the Pricing of Average-Rate Deriva-
tives," mimeo, University of Maryland.
Kendall, M. and A. Stuart, 1977, The Advanced Theory of Statistics, Volume 1, Macmillan, New
York.
Longsta , Francis, 1989, \A Nonlinear General Equilibrium Model of the Term Structure of In-
terest Rates," Journal of Financial Economics 23, 195{224.
Longsta , Francis, and Eduardo Schwartz, 1992, \Interest Rate Volatility and the Term Structure:
A Two-Factor General Equilibrium Model," Journal of Finance 47, 1259-1282.
Lukacs, Eugene, 1960, Characteristic Functions, Charles Grin and Company, London.
Merton, Robert, 1973, \Theory of Rational Option Pricing," Bell Journal of Economics 4, 141-183.

Nachman, David, 1988, \Spanning and Completeness with Options," Review of Financial Studies
1, 311-328.
Ross, Stephen, 1976, \Options and Eciency," Quarterly Journal of Economics 90, 75-89.
Scott, Louis, 1997, \Pricing Stock Options in a Jump-Di usion Model with Stochastic Volatility
and Interest Rates: Application of Fourier Inversion Methods," Mathematical Finance 7, 413-426.

Shephard, Neil, 1991, \From Characteristic Function to a Distribution Function: A Simple Frame-
work for Theory," Econometric Theory 7, 519-529.
Stein, Elias, and Jeremy Stein, 1991, \Stock Price Distributions with Stochastic Volatility: An
Analytic Approach," Review of Financial Studies 4, 727-752.
Stulz, Rene, 1982, \Options on the Minimum or the Maximum of Two Risky Assets," Journal of
Financial Economics 10, 161{185.
Zhang, Peter, 1998, Exotic Options, 2nd Edition, World Scienti c Publishing, New Jersey.

32