Download as pdf or txt
Download as pdf or txt
You are on page 1of 40

Stoptions: Representations and

Applications

Peter Carr pc73@nyu.edu

Initial version: January 9, 2021


Current version: August 20, 2021
File reference: Stoptions.tex

Abstract
We introduce a new derivative security called a stoption. After paying an upfront premium, the
owner of a stoption accrues realized price changes in some underlying security until the exposure is
stopped by the owner. Upon stopping, the reward is the sum of all of the previous price changes plus
a deterministic amount which can vary with the stopping time. Stoptions are finite-lived and hence
must be stopped at or before a fixed maturity date. We propose a particular discrete-time proba-
bilistic model for the underlying’s price changes and then determine the optimal stopping strategy
and stoption premium for that model in closed-form. We also present an application to DVA (debit
valuation adjustment) under full collateralization.

I am grateful to Jerome Benveniste, Umberto Cherubini, Doug Costa, J.P. Delavin, Mike Lipkin,
Federico Maglione, Harvey Stein, Lorenzo Torricelli, and Le Zhu for helpful conversations. They are not
responsible for any errors.

Electronic copy available at: https://ssrn.com/abstract=3929583


1 Introduction
The simplest derivative securities have payoffs that are European-style and path-independent, eg. vanilla
calls and puts. Next up in complexity are payoffs which are European-style but path-dependent eg.
lookback or Asian options. The next level of complexity are payoffs which are Bermudan-style, but path-
independent eg Bermudan swaptions. An even greater level of complexity arises when payoffs are both
Bermudan-style and path-dependent. In this paper, we introduce a naturally arising new derivative security
whose payoff is both Bermudan-style and path-dependent. We call this derivative security a stoption. The
name arises because the owner has optionality over when to stop an exposure to price changes in some
underlying reference security.
After paying an upfront premium, the owner of a stoption accrues realized price changes in the under-
lying reference security until the exposure is stopped by the owner. Upon stopping, the reward is the sum
of all of the previous price changes in the underlying plus a deterministic amount called a floor. The floor
received upon stopping can vary with the stopping time selected by the owner of the stoption. The floor
can be regarded as a replacement for the just realized price change that would have accrued had the owner
not stopped, but rather continued. Stoptions are finite-lived and hence must be stopped at or before a
fixed maturity date.
In this paper, we assume that the spot price s of the underlying security is real-valued. The payoff of
a stoption is a function of the stoption’s term eg. n = 252 business days, the monitoring frequency of its
underlying, eg. daily, an n vector of future random spot prices s ∈ Rn and an n− vector of deterministic
floors a ∈ Rn . If the owner stops at time i = 1, . . . n, then the reward received at time i is simply
si−1 − s0 + ai . Thus by stopping at time i, the stoption owner replaces the last price change si − si−1 which
would have accrued by continuing with a deterministic floor, ai ∈ R instead.
The valuation of a stoption is an example of an optimal stopping problem. The main contribution
of this paper is to propose a particular probabilistic model under which this optimal stopping problem
can be solved in closed-form. In particular, we assume that under a risk-neutral measure, the underlying
security’s price undergoes a logistic random walk. Hence, the n price changes in the underlying security
over the stoption’s life are IID logistically distributed. We also assume zero carrying costs so that the
risk-neutral mean of each underlying security price change is zero. We let b > 0 denote the assumed
known scale parameter of each underlying security price change. In this probabilistic setting, we introduce
ideas from abstract algebra to provide a closed-form solution to the optimal stopping problem determining
the stoption’s initial premium. The resulting valuation formula is particularly simple, relying only on the
ability to calculate the exponential and its inverse. Moreover, we provide explicit expressions for the n
critical price changes which determine the optimal exercise strategy.
The paper contains several alternative representations of the stoption’s initial premium in our model.
For example, linear homogeneity is used to provide a representation in terms of greeks, i.e partial derivatives
w.r.t. floors and the scale parameter of the logistic model. Convex duality is also used to provide a two
term decomposition of the premium isolating the impact from adapting the stopping time to the path of
the underlying. Convex duality also provides a different two term decomposition which we interpret as a
generalized form of put-call-parity.
An overview of this paper is as follows. The next section assumes no frictions, no arbitrage, and no
carrying costs, so that the stoption’s underlying spot price process s is a real-valued martingale under
some risk-neutral probability measure. It then presents a probabilistic representation of the stoption’s
initial premium. The following section gives restrictions on the stoption’s initial premium in this general
martingale setting. We then further assume in the next section that the underlying security’s price changes

Electronic copy available at: https://ssrn.com/abstract=3929583


are IID. We show that the initial premium of an n-period stoption is the result of iterating a single-period
married put’s value function on itself n − 1 times. To solve this iteration in closed form we further assume
in the next section that the underlying security’s price changes are IID logistic with zero mean and known
scale. We show that a novel algebraic approach then allows the stoption’s initial premium and critical price
change vector to be determined in closed-form. We also present a semi-static hedging strategy involving
rolling single-period vanilla options written on the same underlying security. In the following section, we
use the convexity of the stoption’s premium in its floors to represent the stoption premium via convex
duality. In the following section, we explore the financial implications of this representation in our setting.
In the next section, we consider a special case of a stoption contract where the floors are all non-positive
and the initial stoption premium vanishes. In the following section, we present an analogy between first
order conditions arising in the optimizations which respectively generate the stoption premium via convex
duality and the gain version of the CAPM when all betas are positive. In the next section, we present
an application to DVA (debit valuation adjustment) under a full collateralization agreement. In the
subsequent section, we consider replacing the maximum in a stoption’s payoff with a minimum and draw a
connection to Helmholtz free-energy. In the penultimate section, we show how the payoff, the model, and
the closed-form pricing formula change when we allow a non-zero constant riskfree interest rate. In the
final section, we summarize the paper and present suggestions for future research. An appendix considers
the stoption premium in the logistic model as a generating function.

2 Probabilistic Representation of Stoption Premium


In this section, we just assume no frictions, no arbitrage, and no carrying costs for the stoption and its
underlying security. We also assume that the price of the underlying is real-valued. Our assumptions imply
the existence of a risk-neutral probability measure Q such that the underlying spot price s is a real-valued
local martingale. We further assume that s is a Q martingale. Since the stoption’s monitoring frequency
of its underlying security price s is discrete, the supporting time set for the Q martingale s is also the same
discrete time set. In this section, we give a probabilistic representation of the stoption’s initial premium
as the solution to a discrete-time optimal stopping problem.
The stoption is finite-lived so let n be a finite natural number describing the number of exercise
opportunities embedded in the stoption. Let a ∈ Rn be the deterministic n vector of floors. The monitoring
frequency, the reference name, the term n, and the floor vector a would all be in the term sheet of the
stoption.
Recall that with daily monitoring, if the stoption owner exercises on day i = 1, . . . n, then the reward
received on day i is the prior aggregate move plus that day’s floor, i.e. si−1 − s0 + ai . Let mn (a) be a
real-valued function of the floor vector a which gives the initial market price of the stoption that we wish
to determine. Let N ∈ [1, 2, . . . n] be a random stopping time describing the exercise time of the stoption.
N is a categorical random variable, also called multinoulli. Then the probabilistic representation of the
stoption premium mn (a) for a given finite term n ≥ 1 and floor vector a ∈ Rn is:
max
msn (a) =N ∈[1,2,...,n] E0Q [sN −1 − s0 + aN ] . (1)

If n = 1, then the probabilistic representation (1) collapses to the first floor a1 , which is received at time
1:
ms1 (a1 ) = a1 . (2)

Electronic copy available at: https://ssrn.com/abstract=3929583


Thus a stoption can be interpreted as a multi-period generalization of a bond. If n = 2, then the proba-
bilistic representation (1) simplifies to:

ms2 (a) = E0Q [a1 ∨ (s1 − s0 + a2 )]. (3)

Thus, the stoption owner can exercise at time 1 and receive a1 or else continue at time 1 and receive
s1 − s0 + a2 at time 2. Comparing to n = 1, we see that optionality has been added to the payoff.
If n = 3, then the probabilistic representation (1) simplifies to:

ms3 (a) = E0Q [a1 ∨ (s1 − s0 + a2 ) ∨ (s2 − s0 + a3 )]. (4)

Thus, the stoption owner can exercise at time 1 and receive a1 or else continue to time 2. If the stoption
is not exercised at time 1, then it can be exercised at time 2 for s1 − s0 + a2 or else the owner can continue
to time 3. Since time 3 is the maturity date, a previously unexercised stoption must be exercised and the
reward received then is s2 − s0 + a3 . Comparing this 3− period stoption payoff to the 2−period stoption
payoff, we see that optionality has again been added.
For any finite n = 1, 2, . . ., if the n−period stoption is not exercised early, then the final payoff sn−1 −
s0 +an received at time n has the same form as any intermediate payoff si−1 −s0 +ai arising from exercising
early at time i < n.
Suppose that the stoption payoff is altered by replacing the maximum in (1) by a minimum:
min
m− Q
n (a) ≡N ∈[1,2,...,n] E0 [sN −1 − s0 + aN ] . (5)

Now:
min
m−
n (a) ≡ N ∈[1,2,...,n] E0Q {(−1)(−1) [sN −1 − s0 + aN ]}
max
= − N ∈[1,2,...,n] E0Q [s0 − sN −1 − aN ]
= −m−s (−a). (6)

In words, the value of the min payoff can be obtained from the value of the max payoff by negating the
premium, the underlying, and the floors of the latter. We will focus mainly on the max payoff in this
paper. However, the min payoff has an interesting connection to Helmholtz free-energy, as we will show.

3 Model-Free Results for a Stoption’s Premium


In this section, we present some restrictions on the initial premium of the stoption arising from just the
assumptions of the last subsection. In the next section, we strengthen the probabilistic structure to obtain
a unique solution for the optimal stopping strategy and the stoption premium.
In general, a stoption’s payoff is random both in size and in timing. However, we now show that the
floors can be chosen so that the payoff is random only in size but occurs at a known time. It will follow
that a stoption can be regarded as a generalization of a portfolio of cash and its risky underlying security.
Suppose that an investor wishes to receive a deterministic payoff ci at a future deterministic time i,
where i = 1, . . . , n. Let Dij be a Kronecker delta function:
(
1 if j = i
Dij ≡ (7)
0 if j 6= i, i, j = 1, . . . , n.

Electronic copy available at: https://ssrn.com/abstract=3929583


There are at least two ways the Kronecker delta function can be used to sift through a sequence c1 , . . . , cn
and pick out the j-th element:
n
X
ci = cj Dij
j=1
max
ci = j∈[1,2,...,n] [cj + ln Dij ]. (8)

The first equation in (8) is called the sifting property of the Kronecker delta function. The second equation
in (8) is the max-plus version of this sifting property. There is a third way to sift ci out of the sequence,
{cj , j = 1, . . . , n}, which is closely related to the second equation in (8). The third way is to set the floors
of a stoption aj = cj + ln Dij for j = 1, . . . , n. In this case, the optimal stopping time N ∗ reduces to the
deterministic number i. As a result, the stoption’s payoff will reduce to the random amount si−1 − s0 + ci
received at the future deterministic time i. The value at time i−1 of this stoption payoff is also si−1 −s0 +ci .
This random value is matched by the payoff at time i − 1 from a static portfolio holding one share of the
stoption’s risky underlying security, and also keeping −s0 + ci in cash. As a result, the arbitrage-free initial
value of the stoption will reduce to the initial price of this replicating portfolio, which is ci :
max
N ∈[1,2,...,n] E0Q [sN −1 − s0 + cN + ln DiN ] = E0Q [si−1 − s0 + ci ] = ci . (9)

We next derive a useful two-term decomposition of the stoption premium mn (a) given in (1). Letting
N ∗ denote the optimal stopping time, the stoption premium is:

mn (a) = E0Q [sN ∗ −1 − s0 + aN ∗ ] = E0Q [sN ∗ −1 − s0 ] + E0Q aN ∗ , (10)

by the linearity of the expectation operator. Conditioning on the realization of N ∗ in the last term in (10),
the law of total probability implies:
n
X
mn (a) = E0Q [sN ∗ −1 − s0 ] + p∗i ai , (11)
i=1

where for i = 1, . . . , n, p∗i denotes the probability that N ∗ = i. Suppose that we subtract and add sN ∗
inside the expectation on the RHS of (11):
n
X n
X
mn (a) = E0Q [sN ∗ −1 − sN ∗ + sN ∗ − s0 ] + p∗i ai = p∗i ai − E0Q [sN ∗ − sN ∗ −1 ] , (12)
i=1 i=1

since E0Q [sN ∗ − s0 ] = 0, from the optional stopping theorem. The last term in (12) need not vanish even
though s is a Q martingale. Roughly speaking, if exercising the stoption is more likely after a drop than
a rise, then −E0Q [sN ∗ − sN ∗ −1 ] > 0. The more volatile are the price changes in the underlying, the more
positive is this last term.
n
p∗i ai on the RHS of (12), let m̃∗n (a) be an expected value calculation as
P
To obtain intuition on the sum
i=1
in (10), except that the optimal stopping time N ∗ is replaced with an independent stopping time I taking
values in the interval [1, 2, . . . , n]. The optimal stopping time N ∗ depends on the filtration generated by
the s path, but by definition, the stopping time I is independent of this filtration. We assume that the

Electronic copy available at: https://ssrn.com/abstract=3929583


independent stopping time has the same probability vector p∗ as the optimal stopping time N ∗ . For any
finite n = 1, 2 . . . , m̃∗n is calculated as:
n
X
m̃∗n (a) = E0Q [sI−1 − s0 ] + E0Q aI = p∗i ai , (13)
i=1

I1
P
since E0Q [sI−1 − s0 ] = E0Q [ [si − si−1 ] = 0 by Wald’s identity and the martingality of s. Comparing the
i=1
RHS’s of (12) and (13), we see that the use of the optimal stopping time N ∗ instead of an independent
stopping time I causes mn (a) to exceed m̃∗n (a) by an adaption premium −E0Q [sN ∗ − sN ∗ −1 ], which must
be non-negative.
Suppose that we wish to once differentiate the RHS of (11) w.r.t. each floor ai . If the underlying Q
martingale s has a continuous state space, then it is reasonable to assume that mn (a) is differentiable
n
p∗i ai , in (11), but they are also implicitly
P
in each floor ai . The floors ai appear explicitly in the sum
i=1
in the optimized probabilities p∗i . The floors ai are also implicitly in the optimal stopping time N ∗ and
hence in sN ∗ −1 in the adaption premium E0Q [sN ∗ −1 − s0 ]. Nonetheless, by the envelope theorem, one
can ignore these implicit dependencies. when differentiating once w.r.t. the parameters ai . Accordingly,
differentiating (11) w.r.t. ai implies:

mn (a) = p∗i , i = 1, . . . , n. (14)
∂ai
In words, (14) says that the risk-neutral probability p∗i that the optimal stopping time N ∗ = i is just the
slope of the stoption premium mn (a) in the i−th floor ai , for i = 1, . . . , n. Hence, the gradient of mn in a
is a risk-neutral probability vector p∗ .
Since p∗i > 0, mn (a) is increasing in each ai . Moreover, since p∗i < 1, we also have that each slope of
mn (a) in each ai . is also bounded above by one. Since the probability p∗i that N ∗ = i increases with each
ai , mn (a) is also convex in each ai .
Since addition distributes over both expected value and the maximum, shifting each floor ai ∈ R by a
common amount δ ∈ R shifts the premium by the same amount:
n
_
mn (a + δ1) = E0Q [sN −1 − s0 + aN + δ] = δ + mn (a). (15)
N =1

Consider the special case of (15) when the floor vector is constant at some level a ∈ R, i.e. a = a1. We
can then set δ = −a in (15), so that after re-arranging, (15) becomes:
mn (a1) = a + mn (0), δ ∈ R. (16)
Now, obviously, the elements of the vector a lie between the smallest and largest values:
n
^ n
_
ai 1 ≤ a ≤ ai 1. (17)
i=1 i=1

Recalling that mn is increasing in the floor vector, we have:


n
! n
!
^ _
mn ai 1 ≤ mn (a) ≤ mn ai 1 . (18)
i=1 i=1

Electronic copy available at: https://ssrn.com/abstract=3929583


Using (16) on the far LHS and RHS of (18) implies after re-arranging that:
n
^ n
_
ai ≤ [mn (a) − mn (0)] ≤ ai . (19)
i=1 i=1

In words, a stoption spread priced at mn (a) − mn (0) is bounded above by the highest floor and bounded
below by the lowest floor. The stoption premium is also bounded below by the largest floor:
n
_
mn (a) ≥ ai . (20)
i=1

This lower bound is reached if the underlying security price is constant for n periods. Combining (19) and
(20) implies the following bounds for the stoption premium mn (a):
n n
! n
_ ^ _
ai ∨ ai + mn (0) ≤ mn (a) ≤ ai + mn (0). (21)
i=1 i=1 i=1

Suppose that we re-arrange (15) as:

mn (a) = mn (a + δ1) − δ, (22)

for any natural number n and any δ ∈ R. Setting n = 2 and δ = −a1 in (22) and using (3) implies:

m2 (a) = a1 + m2 (a − a1 1) = a1 + E0Q [0 ∨ (s1 − (s0 + a1 − a2 )]. (23)

By the RHS of (23), the two-period stoption has the same initial value as the first floor a1 plus a one-period
call on the same underlying and struck a1 − a2 dollars above the initial spot s0 . Put-call-parity implies
that:
E0Q [0 ∨ (s1 − (s0 + a1 − a2 )] = s0 − (s0 + a1 − a2 ) + E0Q [(s0 + a1 − a2 − s1 ) ∨ 0]. (24)
Substituting (24) in (23) implies:

m2 (a) = a2 + E0Q [(s0 + a1 − a2 − s1 ) ∨ 0]. (25)

Thus, the two-period stoption has the same initial value as the second floor a2 plus a one-period put on
the same underlying and struck a1 − a2 dollars above the initial spot s0 .
The decompositions (23) and (25) suggest potential stoption replication strategies using bonds and
vanilla options. Since a stoption and a vanilla call both have positive delta, the 2 period stoption has the
same value at time 1 as a portfolio holding a bond worth a1 and a one-period vanilla call struck a1 − a2
dollars above s0 . Since a vanilla put has negative delta, the 2 period stoption has the same value at time
1 as a portfolio holding a bond worth a2 , a zero-cost forward contract paying s1 − s0 at time 1 and the
one-period vanilla put struck a1 − a2 dollars above s0 . The latter portfolio can be called a (one period)
married put.
Later in the paper, we will present a generalization of these two replication results to an n−period
stoption. In a particular model described in the next section, an n−period stoption has the same value at
time 1 as a portfolio of a bond and a one-period vanilla call with an appropriately chosen strike.

Electronic copy available at: https://ssrn.com/abstract=3929583


4 Random Walk and Function Iteration
In this section, we assume that under a risk-neutral measure Q, the spot price s of the stoption’s underlying
security follows a discrete-time random walk with known initial value s0 ∈ R. We will show that the
assumption of independently and identically distributed (IID) increments for s implies that the n−period
stoption premium arises as the result of iterating a single-argument function with itself n − 1 times.
The function being iterated combines a parameter and an argument into the value of a married put.
The function’s parameter is the strike price of the married put, while the function’s argument is the
contemporaneous price of the married put’s underlying security.
Consider the problem of valuing a 1-period married put at time i − 1 with strike price ai ∈ R. The
married put is not written on the stoption’s underlying security price si ∈ R per se, but is instead written
on a security whose price is just a translation of si . Let ui ≡ ai+1 + si − si−1 denote the random value
of the married put’s underlying security when the married put matures at time i. Since si−1 is known at
time i − 1, the married put’s underlying security price ui is just a translation ai+1 − si−1 of si . The married
Q
put’s payoff at time i is ai ∨ ui , where ui has risk-neutral mean ai+1 , i.e. Ei−1 ui = ai+1 . The married
ai Q
put value function Mi−1 (ai+1 ) ≡ Ei−1 [ai ∨ (ai+1 + si − si−1 )] relates the married put’s value Mi−1 at time
i − 1 to the married put’s strike price ai ∈ R and to the contemporaneous price ai+1 ∈ R of the married
ai
put’s underlying security. The superscript ai of Mi−1 is the strike price treated as a parameter, while the
ai
argument ai+1 of Mi−1 (ai+1 ) is the contemporaneous price of the married put’s underlying security.
Since the stoption’s underlying security price process s is a time-homogeneous random walk, the married
put value function has no direct dependence on time given the strike price ai ∈ R and the contemporaneous
value ai+1 ∈ R of the married put’s underlying security. As a result, we can drop the time subscript on
the married put value function:

M ai (ai+1 ) ≡ Ei−1
Q
[ai ∨ (ai+1 + si − si−1 )], (26)

where for i = 1, . . . , n, the increments si − si−1 are IID with zero Q mean.
(n−i)
We now value an n-period stoption using dynamic programming. Let CVi be the continuation
value of the n-period stoption at day i, for i = 0, 1, . . . , n, assuming no prior exercise at periods 0, 1, . . . , i.
At day n − 1, the stoption must be exercised at period n for a payoff of an + sn−1 − s0 . Since this payoff
is known at period n − 1, we set:
(1)
CVn−1 = an + sn−1 − s0 . (27)
Stepping back to period n − 2, the continuation value then is:
(2) Q (1)
Q
CVn−2 = En−2 [(an−1 + sn−2 − s0 ) ∨ CVn−1 ] = En−2 [(an−1 + sn−2 − s0 ) ∨ (an + sn−1 − s0 )], (28)

from (27). Since sn−1 − s0 = sn−1 − sn−2 + sn−2 − s0 , we can factor out sn−2 − s0 :
(2) Q
CVn−2 = sn−2 − s0 + En−2 [an−1 ∨ (an + sn−1 − sn−2 )] = sn−2 − s0 + M an−1 (an ), (29)

from (26).
Stepping back to period n − 3, the continuation value then is:
(3) Q (2)
CVn−3 = En−3 [(an−2 + sn−3 − s0 ) ∨ CVn−2 ]
Q
= En−3 [(an−2 + sn−3 − s0 ) ∨ (sn−2 − s0 + M an−1 (an ))], (30)

Electronic copy available at: https://ssrn.com/abstract=3929583


from (29). Factoring out sn−3 − s0 from the expected maximum, the continuation value at period n − 3 is:
(3) Q
CVn−3 = sn−3 − s0 + En−3 [an−2 ∨ (sn−2 − sn−3 + M an−1 (an ))]
= sn−3 − s0 + M an−2 (M an−1 (an )), (31)
(3)
from (26). Thus, the continuation value CVn−3 at period n − 3 is just a known translation sn−3 − s0 of an
iterated function M an−2 (M an−1 (an )). The function M ai (ai+1 ) being iterated is the value of a 1-day married
put with strike ai and underlying ui = ai+1 + si − si−1 with Q mean ai+1 . Using induction, it can be shown
that the n-period stoption value is also described by an n − 1-fold function iteration:
(n) a
CV0 = M2a1 (M2a2 (. . . M2 n−1 (an ) . . .)). (32)

To summarize the results of this section, the initial premium for an n-period stoption written on the
increments of a discrete-time random walk can be determined by iterating the value of a single-period
married put written on a translation of the stoption’s underlying security price. To uniquely determine
the n-period stoption’s initial premium, the modeller need only pick some zero-mean distribution with real
support to describe the law of each increment si+1 − si . Ideally, this choice of law leads to a closed-form
formula for the single-period married put value function, but this is not necessary. Initializing the recursion
for M ai (ai+1 ) with parameter an−1 and with argument an , iterating the single-period married put value
(n)
function n − 1 times results in the initial premium CV0 given in (32) for an n-period stoption.
For some functions (eg. affine), the result after repeatedly iterating can be expressed in closed-form.
The set of functions with this property is quite small. Fortunately, we are aware of a distribution with
zero mean and with real support which leads to closed-form formulas for both the married put value and
the result of iterating this married put value function on itself n − 1 times. We introduce this distribution
in the next section.

5 The Logistic Model Solution for Initial Stoption Premium


In this section and for the rest of the paper, we further assume that the underlying security’s price changes
si+1 − si are IID logistically distributed with zero mean and known scale parameter b > 0. This logistic
option pricing model (LOPM) allows us to give a novel algebraic representation of the stoption’s initial
premium. The LOPM also determines the risk-neutral distribution of the optimal stopping time N ∗ ,
which is a categorical random variable. The probability mass function of N ∗ will be shown to be a Gibbs
distribution, with each Gibb’s probability p∗i ∈ (0, 1), i = 1, . . . , n distinguished from the other probabilities
by its corresponding floor ai ∈ R.
The main idea underlying the LOPM is to change from a classical arithmetic based on ordinary addition
and ordinary multiplication over non-negative reals to a new isomorphic arithmetic over reals with non-
standard addition and multiplication operations. A two-period stoption has the same value as a European-
style option and this common value will be seen as the outcome of applying a new binary operation. The
n−period stoption premium will also be seen as a new type of repeated sum. See Burgin and Czachor[4]
for an overview of alternative arithmetics and applications to physics and psychology. Here, we will apply
the idea to derivative security pricing in finance and motivate the approach using a traditional probabilistic
framework.

Electronic copy available at: https://ssrn.com/abstract=3929583


A continuous real-valued random variable X is said to be logistically distributed under a risk-neutral
probability measure Q if its cumulative distribution function (CDF) has the form:
x−µ −1
 
Q{X ≤ x} = 1 + e− b , x ∈ R, (33)

for some real µ and for some positive real b. When X is a logistically distributed random variable, the
real parameter µ is the risk-neutral mean of X and the positive parameter b is a scale parameter. The
scale parameter b > 0 is translation-invariant, i.e. if X has scale parameter b > 0, then so does the logistic
random variable X − δ for any constant translation δ ∈ R. Moreover, for any λ > 0, the logistic random
variable X−δ
λ
has scale parameter λb . When the logistic random variable X has units eg. dollars, then µ
and b must have the same units and the standard logistic random variable X−µ b
is dimensionless with mean
zero and scale one.
The scale parameter b is not the standard deviation of X per se, but b is positively proportional to the
standard deviation σ: √ √
3 3p
b= σ= E[(X − µ)2 ]. (34)
π π
The scale parameter b is also positively proportional to the expected payoff from a call written on X and
struck at its mean µ:
1
b= E[(X − µ)+ ]. (35)
ln 2
If X1 and X2 are IID logistic with common mean µ ∈ R and common scale b > 0, then:

b = E[(X1 − X2 )+ ] = E[(X2 − X1 )+ ]. (36)

The 3 formulas (34),(35) and (36) for b get progressively simpler and they all capture the notion that b is
a positive measure of the spread or width of the probability density function of the logistically distributed
random variable X. In a financial context, we refer to b as the bewilderment about each future spot price
change si − si−1 , i = 1, . . . , n.
The support of the logistic distribution is the whole real line R. Since the realization for a spot price
change can be arbitrarily negative, the spot price level can be negative. As a result, the LOPM should be
considered as an alternative to the Bachelier[2] option pricing model, as opposed to the Black-Scholes[3]
model. When the underlying of a European option is a swap value, an interest rate, a temperature, or a
crude oil futures price, models with real-valued underlyings such as the Bachelier model or the LOPM are
more appropriate than the Black-Scholes model.
Let Z1 , . . . , Zn denote a set of n IID standard logistic random variables each with mean zero and scale
one. The LOPM is the assumption that under a risk-neutral measure Q, each underlying security’s price
change, si+1 − si = bZi+1 for i=0, 1, . . . , n-1. It then follows that the underlying security’s price changes
are IID logistic, all with zero mean and known scale parameter b > 0.
We indicate the price of an n−period stoption in the LOPM model by:
"N −1 #
max X
mn (a, b) ≡N ∈[1,2,...,n] E0Q bZi + aN . (37)
i=1

A 1-period stoption is priced at a1 in any model including the LOPM. Hence, we now consider the pricing

Electronic copy available at: https://ssrn.com/abstract=3929583


of a 2-period stoption in the LOPM. Setting n = 1 in (37):

m2 (a, b) = E0Q [a1 ∨ (bZ1 + a2 )]


= E0Q {(bZ1 + a2 ) + [(a1 − bZ1 − a2 ) ∨ 0]}
= a2 + E0Q [−(a2 − a1 + bZ1 ) ∨ 0]
= a2 + E0Q [(−X) ∨ 0], (38)

where X is logistically distributed with mean a2 − a1 and scale b > 0. Now for any continuous random
variable Y , we have representations in terms of Dirac delta and Heaviside functions:
Z 0 Z 0
(−Y ) ∨ 0 = (−x)δ(Y − x)dx = 1(Y < x)dx, (39)
−∞ −∞

using integration by parts. When Y has a finite lower partial mean under Q, one also has:
Z 0 Z 0
Q Q
E0 [(−Y ) ∨ 0] = E0 1(Y < x)dx = Q{Y < x}dx. (40)
−∞ −∞

Hence, when Y is the logistic random variable X with mean a2 − a1 and scale b > 0, the LOPM value of
the payoff (−X) ∨ 0 in (38) is:
Z 0 
x−(a2 −a1 ) −1

Q
E0 [(−X) ∨ 0] = 1 + e− b dx
−∞
Z 0  x−(a2 −a1 )

= db ln 1 + e b

−∞
 0
x−(a2 −a1 )

= b ln 1 + e b

−∞
 a1 −a2

= b ln 1 + e b . (41)

Hence from (38), we have:


 a1 −a2
  a2
  a1 −a2
  a1 a2

m2 (a, b) = a2 + b ln 1 + e b = b ln e b + b ln 1 + e b = b ln e + e
b b . (42)

In words, the RHS of (42) indicates


a1
that the real-valued floors a1 and a2 are mapped via exponentiation
a2
into the positive quantities e and e b respectively. It will prove useful to refer to these positive quantities
b

as un-normalized Gibbs probabilities. From the RHS of (42), the two un-normalized Gibbs probabilities
are then added and the inverse map b ln · is applied to price the two period stoption in the LOPM.

5.1 The Log-Plus Semi-field


In this section, we show how abstract algebra can be used to further simplify the LOPM value of a 2-period
stoption. Recall that a binary operation ⊕ is said to be associative over a set S if for any 3 values a1 , a2 ,
and a3 in the set, we have:
(a1 ⊕ a2 ) ⊕ a3 = a1 ⊕ (a2 ⊕ a3 ). (43)

10

Electronic copy available at: https://ssrn.com/abstract=3929583


Ordinary addition over positive reals is an example of an associative binary operation. Aczel[1] proved
that a binary operation ⊕ is associative over a set S if and only if there exists an increasing function G(a)
of a, called a generator such that for any a1 , a2 ∈ S:

a1 ⊕ a2 = G G−1 (a1 ) + G−1 (a2 ) .



(44)

Roughly speaking, all associative binary operations ⊕ inherit their associativity from ordinary addition +.
Borrowing terminology from Pap[9], we refer to a1 ⊕ a2 as a pseudo-sum of a1 and a2 . Comparing (42)
with (44), we have G(a) = b ln g, for g > 0 and b > 0. For each fixed b > 0, we let ⊕b be a binary operation
arising from the addition of positive real numbers g > 0 when G(g) = b ln g, g > 0 is the generator. Hence,
we can interpret the LOPM value of a 2-period stoption as a pseudo-sum of the two floors:
 a1 a2

m2 (a, b) = a1 ⊕b a2 ≡ b ln e b + e b , a1 , a2 ∈ R, b > 0. (45)

We regard the middle expression a1 ⊕b a2 as not just a suggestive notation for the log-sum-exponential on
the RHS of (45), but also as a simpler pricing formula for the LOPM stoption premium m2 (a; b). From
Aczel’s characterization, the binary operation ⊕b is associative over the real line R ≡ (−∞, ∞), and it is
also clearly commutative:
 a2 a1
  a1 a2

a2 ⊕b a1 = b ln e b + e b = b ln e b + e b = a1 ⊕b a2 = m2 (a, b). (46)

It follows that over R, the algebraic structure (R; ⊕b ) is a commutative semi-group.


This simple algebraic structure can be extended. Setting a2 = −∞ in (45) implies:
 a1 −∞

a1 ⊕b (−∞) = b ln e + e b b = a1 . (47)

Thus, −∞ is the neutral element for ⊕b . Over the set [−∞, ∞), the algebraic structure ([−∞, ∞); ⊕, −∞)
is a commutative monoid.
Now, ([−∞, ∞); +, 0) is a commutative group. Moreover, on [−∞, ∞), the binary operation + right
and left distributes over ⊕b for all b > 0:

(a1 ⊕b a2 ) + a3 = (a1 + a3 ) ⊕b (a2 + a3 ) a3 + (a1 ⊕b a2 ) = (a3 + a1 ) ⊕b (a3 + a2 ). (48)

It follows that the algebraic structure ([−∞, ∞); ⊕b , −∞; +, 0) is a commutative semi-field, for any fixed
b > 0. This semi-field is isomorphic to the familiar semi-field ([0, ∞); +, 0; ×, 1). For any fixed b > 0, the
generator G(g) = b ln g, g ≥ 0 maps the commutative semi-field ([0, ∞); +, 0; ×, 1) to the commutative semi-
field ([−∞, ∞); ⊕b , −∞; +, 0). Borrowing terminology from McEneaney[8], we call ([−∞, ∞); ⊕b , −∞; +, 0)
the log-plus semi-field.

5.2 Solution for the Initial Premium of an n−Period Stoption


In this subsection, we use the binary operation ⊕b to give a simple formula for the LOPM initial value of
an n−period stoption. We also describe the additional mathematical properties that this LOPM formula
has over the general value of a stoption written on the increments of any Q martingale.

11

Electronic copy available at: https://ssrn.com/abstract=3929583


Recall that (45) is the LOPM stoption premium when n = 2. When n = 3, the reward for stopping at
time 3 is s2 − s0 + a3 = bZ1 + bZ2 + a3 , while the reward for stopping at time 2 is s1 − s0 + a2 = bZ1 + a2 .
Thus, under the LOPM, the probabilistic representation (4) becomes:

m3 (a, b) = E0Q [a1 ∨ (bZ1 + a2 ) ∨ (bZ1 + bZ2 + a3 )] = E0Q [a1 ∨ (bZ1 + (a2 ∨ (bZ2 + a3 )))], (49)

since + distributes over ∨. Using the law of iterated expectations:

m3 (a, b) = E0Q E1Q [a1 ∨ (bZ1 + (a2 ∨ (bZ2 + a3 )))] = E0Q [a1 ∨ (bZ1 + E1Q [a2 ∨ (bZ2 + a3 )])], (50)

since Z1 is known at time 1. Now E1Q [a2 ∨ (bZ2 + a3 )])] is the LOPM value at time 1 of a 2-period stoption
with floors a2 and a3 . Using our algebraic representation:

m3 (a, b) = E0Q [a1 ∨ (bZ1 + (a2 ⊕b a3 ))]. (51)

Now, the RHS of (51) is the LOPM initial value of a 2-period stoption with floors a1 and a2 ⊕b a3 . From
our algebraic representation:

m3 (a, b) = a1 ⊕b (a2 ⊕b a3 ) = a1 ⊕b a2 ⊕b a3 , (52)

by the associativity of ⊕. We conclude that the LOPM initial value of a 3-period stoption is:
3  a1 
M a2 a3
m3 (a, b) = ai ≡ a1 ⊕b a2 ⊕b a3 = b ln e + e + e
b b b . (53)
b
i=1

This is the natural extension of the LOPM initial value of a 2-period stoption:
2  a1 
M a2
m2 (a, b) = ai ≡ a1 ⊕b a2 ≡ b ln e b + e b . (54)
b
i=1

To illustrate these formulas, suppose that a1 = ln 3, a2 = ln 4, a3 = ln 12, and b = 1/2. The 1-period
premium is just the time 1 payoff a1 = ln 3. Adding optionality to the payoff raises the value to the
following LOPM 2-period stoption premium:
1
m2 ([ln 3 ln 4], 1/2) = ln 32 + 42 2 = ln 5. (55)

Adding further optionality to the payoff raises the value to the following LOPM 3-period stoption premium:
1
m3 ([ln 3 ln 4 ln 12], 1/2) = ln 52 + 122 2 = ln 13. (56)

Dynamic programming can be used to determine the LOPM initial premium of the n−period stoption.
(j)
Recall that CVn−j denote the stoption’s continuation value at future time n−j, conditional on not exercising
at times 1 to n − j inclusive. The subscript n − j on CV indicates the discrete calendar time index, while
the superscript (j) on CVn−j indicates the number of remaining exercise opportunities. We begin the
backward recursion at the discrete calendar time n − 1. The stoption must be exercised at time n, so there
is one opportunity to exercise remaining. The payoff at time n will be sn−1 − s0 + an which is known at
time n − 1. As a result:
(1) Q
CVn−1 = En−1 {sn−1 − s0 + an } = sn−1 − s0 + an . (57)

12

Electronic copy available at: https://ssrn.com/abstract=3929583


This is the continuation value at time n − 1. We step back one period to time n − 2. The stoption can
either be exercised at time n − 1 for a payoff of sn−2 − s0 + an−1 or the owner can continue. The LOPM
stoption value at time n − 2 is the conditional expectation of the larger of these two choices:
(2)
Q (1)
CVn−2 = En−2 {(sn−2 − s0 + an−1 ) ∨ CVn−1 }
Q
= En−2 {(sn−2 − s0 + an−1 ) ∨ (sn−1 − s0 + an )}, (58)

from (57). Since sn−1 − s0 = sn−2 − s0 + bZn−1 , we can factor sn−2 − s0 out on the left:
(2) Q
CVn−2 = sn−2 − s0 + En−2 {an−1 ∨ (bZn−1 + an )}
= sn−2 − s0 + (an−1 ⊕b an ). (59)

This is the continuation value at time n − 2. We step back one period to time n − 3. The stoption can
(2)
either be exercised at time n − 2 for a payoff of sn−3 − s0 + an−2 or the owner can continue and get CVn−2 .
The LOPM stoption value at time n − 3 is:
Q (3) (2)
CVn−3 = En−3 {(sn−3 − s0 + an−2 ) ∨ CVn−2 }
Q
= En−3 {(sn−3 − s0 + an−2 ) ∨ (sn−2 − s0 + (an−1 ⊕b an )), (60)

from (59). Since sn−2 − s0 = sn−3 − s0 + bZn−2 , we can factor sn−3 − s0 out on the left:
(3) Q
CVn−3 = (sn−3 − s0 ) + En−3 {an−2 ∨ (bZn−2 + (an−1 ⊕b an ))}
= sn−3 − s0 + (an−2 ⊕b (an−1 ⊕b an ))
= sn−3 − s0 + (an−2 ⊕b an−1 ⊕b an ), (61)

by the associativity of ⊕b . Continuing in this manner, we can step back to time 0:


n  a1 
M a2 an
(n)
mn (a, b) = CV0 = ai ≡ a1 ⊕b a2 ⊕b . . . ⊕b an = b ln e b + e b + . . . + e b , (62)
b
i=1

for ai ∈ R, i = 1, . . . , n, b > 0. Thus, we have succeeded in giving a closed-form solution to the optimal
stopping problem determining the initial premium of an n−period stoption. The LOPM initial premium
of an n−period stoption is just the n−term log-sum-exponential function. This solution clearly depends
only on the ability to compute the exponential function and its inverse.
Our LOPM stoption pricing formula (62) uses the natural log, but the choice of base e for the logarithm
1
is not required. Since b > 0, we have that β ≡ e b > 1 and that b = 1/ ln β. Expressed in terms of base β,
Mn
the LOPM stoption premium ai simplifies to:
b
i=1

n
M
1
ai = logβ (β a1 + β a2 + . . . + β an ) , ai ∈ R, i = 1, . . . , n, β > 1. (63)
ln β
i=1

As a result, we can also refer to the LOPM scale parameter/bewilderment b > 0 as the base controller.

13

Electronic copy available at: https://ssrn.com/abstract=3929583


Suppose that we increase n by one by extending the n−vector of floors a by any real element an+1 .
n
W n+1
W
Then ai ≤ ai and correspondingly, the shorter-dated stoption has lower LOPM value than the
i=1 i=1
longer-dated one:
n
M n+1
M
ai ≤ ai . (64)
b b
i=1 i=1

The upper bound in (64) is reached if and only if an+1 = −∞. The inequality in (64) is consistent with the
rough statement that real numbers a ∈ R interact with ⊕b and + in the same way that positive numbers
a
g = e b interact with + and ×. Similarly, −∞ interacts with ⊕b and + in the same way that 0 interacts
with + and ×.
Recall that with Dij a Kronecker delta function, setting the stoption’s floors aj = cj + ln Dij , j =
1, . . . , n, causes the stoption’s penultimate value to reduce to the payoff from a static position in cash and
the stoption’s risky underlying security. In the LOPM, (9) becomes:
n
max M
ci =N ∈[1,2,...,n] E0Q [sN −1 − s0 + cN + ln DiN ] = (cj + ln Dij ). (65)
b
j=1

The RHS of (65) can be compared to the sifting property of a Kronecker delta function and its tropical
counterpart:
n
X
Ci = (Cj × Dij )
j=1
max
ci = j∈[1,2,...,n] [cj + ln Dij ]. (66)
The first equation in (66) arises from (65) by switching arithmetic using an exponential generator and
defining Ci = eci . The second equation in (66) arises as a limiting value of (65), as b ↓ 0.
We now describe some of the additional mathematical properties that the LOPM stoption premium
formula (62) has over the general value of a stoption written on the increments of any Q martingale. First,
as the LOPM stoption premium is a commutative pseudo-sum, any reordering/permutation of the floors
preserves value. This invariance need not hold outside of the LOPM. Second, the LOPM pricing formula
n n
M P ai
is computationally inexpensive. From (62), the LOPM stoption premium mn (a, b) = ai = b ln e b ,
b i=1
i=1
which just requires one natural log, ln, and n exponential functions, exp, to evaluate. If all floors vanish,
a = 0, then the exponential functions are no longer needed since:
n
M n
X
mn (0, b) = 0 = b ln 1 = b ln n. (67)
b
i=1 i=1

This simple result describes how the LOPM premium of an at-the-money stoption grows with the number
of exercise opportunities, n.
The imposition of the LOPM allows us to simplify and improve the bounds aon the stoption premium
i
given in (21). To improve the lower bound, let Ā be the simple average of the e b :
n
1 X ai
Ā ≡ e b > 0. (68)
n i=1

14

Electronic copy available at: https://ssrn.com/abstract=3929583


Now b ln · is a concave function so by Jensen’s inequality:
n
1X ai
b ln Ā ≥ b ln(e b ). (69)
n i=1

Equivalently: !
n n
1 X ai 1X
b ln eb ≥ ai . (70)
n i=1 n i=1
As a result, the stoption premium’s lower bound improves to:
n n
M 1X
ai ≥ b ln n + ai . (71)
b n i=1
i=1

This lower bound is reached if and only if all of the floors are equal. Applying a mean preserving spread
Mn
raises the stoption premium ai , while preserving the lower bound on the RHS of (71). Substituting
b
i=1
(71) and (67) in (21) gives the following simple lower and upper bounds on the LOPM stoption premium
M n
ai :
b
i=1 ! !
n n n n
_ 1X M _
ai ∨ b ln n + ai ≤ ai ≤ b ln n + ai . (72)
i=1
n i=1 b
i=1 i=1
n
M
As the scale parameter b ↓ 0, the LOPM stoption premium ai limits to the lower bound on the LHS
b
i=1
of inequality (72), which simplifies:
n
M n
_
lim ai = ai . (73)
b↓0 b
i=1 i=1
n
M
For n = 2, 3, . . . , the LOPM stoption premium ai is increasing in the base controller/bewilderment
b
i=1
b > 0, as we now show. Define an `p norm of a row vector ~g = [g1 g2 . . . gn ] by:

`(~g ; p) ≡ (g1p + g2p + . . . + gnp )1/p .

It is well known, see eg. Raı̈ssouli and Jebri[10], that `(~g ; p) is declining in p. Defining b = 1/p, it follows
n
M
~a
that `(e ; 1/b) is increasing in b. Since ai = ln `(e~a ; 1/b), so it is also increasing in b. We will connect
b
i=1
n
M

∂b
ai > 0 to Shannon entropy later in this paper.
b
i=1
When the n floors ai are all constant at a, then the LOPM stoption premium simplifies to:
n
M
a = a + b ln n, a ∈ R, b > 0, n = 1, 2, . . . (74)
b
i=1

15

Electronic copy available at: https://ssrn.com/abstract=3929583


When the floors grow linearly with time, this pseudo-geometric progression is given by:
n−1
M
(a + i) = (na 0) − (a 0) for a > 0
i=1
= (0 na) − (0 a) for a < 0. (75)
Every ordinary finite series with a closed-form, eg. Gabriel’s staircase would also have an analogous
closed-form.
We now temporarily assume that the stoption has an infinite number of floors. A stoption whose
number of floors is some finite natural number n can meet our assumption by setting floor an+1 = −∞,
setting floor an+2 = −∞, etc. Thus, the finite series of floors is padded by pseudo-zeros to become an
infinite series of floors. For technical reasons, we also introduce a zeroth floor, which can again be −∞.
Thus the stoption premium becomes the following pseudo-series:

M
m∞ (a, b) = aj , (76)
b
j=0

where for j = 0, 1, 2, . . ., each aj ∈ R−∞ and b > 0. The sequence has to eventually decline towards −∞
for the pseudo-series to exist.
We now structure each floor so that the stoption premium m∞ (a, b) becomes what we will call a pseudo-
generating function. To accomplish this, we introduce coefficients cj ∈ R and argument z ∈ R−∞ and set
each floor aj = cj + zj. We call the resulting stoption premium ĉ(z, b):

M
ĉ(z, b) ≡ (cj + zj). (77)
b
j=0

To relate this to a more familiar power series, use the definition of ⊕b :



!
X cj +zj
ĉ(z, b) = b ln e b . (78)
j=0
cj z
Let Cj ≡ e b and let Z ≡ e b :

!
X
ĉ(z, b) = b ln Cj Z j . (79)
j=0

Thus ĉ(z, b) is just b ln · of a generating function of the sequence Cj . As a result, we call ĉ(z, b) a pseudo-
generating function of the sequence cj ∈ R−∞ . The appendix shows how to invert a pseudo-generating
function for a given cj . It also shows that the pseudo-generating function of a pseudo-convolution of two
sequences is just an ordinary sum of the two pseudo-generating functions.
Mn
For each finite n = 1, 2, . . ., the LOPM stoption premium ai is increasing in each ai ∈ R. Recall
b
i=1
from (14) that each first partial derivative of the stoption premium w.r.t. a floor ai is a probability
p∗i ∈ [0, 1]:
ai
n n
!
∗ ∂ M ∂ X ai eb
pi ≡ ai = b ln eb = P n , i = 1, . . . , n. (80)
∂ai b ∂ai ai
i=1 i=1 eb
i=1

16

Electronic copy available at: https://ssrn.com/abstract=3929583


From the RHS of (80), the sum of these partial derivatives is indeed one for each n, i.e.:
n
X
p∗i = 1. (81)
i=1

The n probabilities in (80) describe the Gibbs probability mass function (PMF). As previously indicated
just below (14), each p∗i is the risk-neutral probability that the optimal stopping time N ∗ = i. The
M n
denominator in the RHS of (80) is called the partition function. The LOPM stoption premium ai =
b
i=1
n
P ai
b ln e b arises from applying b ln(·) to this partition function.
i=1
Recall that (15) showed the effect of translating each floor ai by a common shift δ ∈ R. In the LOPM,
this is just the distributivity of + over ⊕b :
n
M n
M
(ai + δ) = δ + ai , δ ∈ R. (82)
b b
i=1 i=1

Now consider scaling each floor ai by a common factor λ > 0 and also scaling b > 0 by the same amount.
Mn
The effect on the LOPM stoption premium ai is to scale it by the same amount:
b
i=1

n
M n
M
λai = λ ai , λ > 0. (83)
λb b
i=1 i=1

Thus, the LOPM stoption premium is linearly homogeneous in the floor vector a and in bewilderment b:

mn (λa; λb) = λmn (a; b), λ > 0. (84)

From Euler’s theorem:


n n n n n n
M X ∂ M ∂ M X
∗ ∂ M
ai = ai ai + b ai = p i ai + b ai . (85)
b
i=1
∂ai b ∂b b
i=1
∂b b
i=1 i=1 i=1 i=1

from (14), where for i = 1, . . . , n, each Gibbs probability p∗i is given in (80). Comparing (85) with (11), we
can conclude that the derivative of the LOPM stoption premium w.r.t. ln b captures the value of having
the probabilities of exercising depend on the price path of s and the floors:
n
∂ M
b ai = E0Q [sN ∗ −1 − s0 ] . (86)
∂b b
i=1

The random variable sN ∗ −1 − s0 is added to aN ∗ to get the reward from exercising the stoption at time
N ∗ . The complement of this random variable is the last price change before exercising sN ∗ − sN ∗ −1 . We
have more to say about this price change in the next subsection.
Differentiating (80) w.r.t. ai implies:
∂ ∗
b pi = p∗i (1 − p∗i ) (87)
∂ai

17

Electronic copy available at: https://ssrn.com/abstract=3929583


n
M
This ordinary differential equation (ODE) is called the logistic ODE. Since p∗i = ∂
∂ai
ai from (14):
b
i=1

n
∂2 M p∗i (1 − p∗i )
ai = . (88)
∂a2i b b
i=1

n
M
Since b > 0 and the numerator of the fraction is positive, (88) confirms that the stoption premium ai
b
i=1
is convex in each ai .
n
M
Since the stoption premium ai is a Fenchel-Legendre transform, it is also convex in the vector
b
i=1
a ∈ Rn . This implies that its Hessian is positive definite. We now use this result to show that the stoption
Mn
premium ai is convex in b > 0. Differentiating (85) w.r.t. b:
b
i=1

n n n n n
∂ M X ∂2 M ∂ M ∂2 M
ai = ai ai + ai + b 2 ai . (89)
∂b b
i=1
∂a i ∂b b ∂b b ∂b b
i=1 i=1 i=1 i=1

n
M

Cancelling ∂b
ai implies:
b
i=1

n n n n n
∂2 M
2
X ∂2 M X ∂2 M
b 2 ai = − ai ai = − aj ai . (90)
∂b b
i=1
∂a i ∂b b
j=1
∂a j ∂b b
i=1 i=1 i=1

Differentiating (85) w.r.t. aj :


n n n n n
∂ M X ∂2 M ∂ M ∂2 M
ai = ai ai + ai + b ai . (91)
∂aj b
i=1
∂ai ∂aj b ∂aj b ∂b∂aj b
i=1 i=1 i=1 i=1

n
M

Cancelling ∂aj
ai implies:
b
i=1

n n n
∂2 M X ∂2 M
b ai = − ai ai (92)
∂b∂aj b
i=1
∂ai ∂aj b
i=1 i=1

Substituting (92) on the RHS of (90):


n n
" n n
#
2 M 2
∂ X X ∂ M
b2 2 ai = − aj − ai ai
∂b b
j=1 i=1
∂ai ∂aj b
i=1 i=1
n X
n 2 n
X ∂ M
= ai aj ai . (93)
j=1 i=1
∂ai ∂aj b
i=1

18

Electronic copy available at: https://ssrn.com/abstract=3929583


n
M
The RHS of (93) is a quadratic form involving the Hessian of ai . Since this Hessian is positive semi-
b
i=1
n
M
∂2
definite, the quadratic form is non-negative. It follows from (93) that ∂b2
ai ≥ 0 so that the stoption
b
i=1
n
M
premium ai is convex in b > 0. Since the constant floor stoption premium in (74) is affine in b, we
b
i=1
know that the convexity in b > 0 is not strict.
Let Mn be a multi-nomial random variable having n outcomes ai and having n probabilities pi ≥ 0, for
n
P
i = 1, . . . , n. Note that while pi = 1, the probabilities pi governing the realization of Mn can differ from
i=1
the Gibbs probabilities p∗i given in (80). For s > 0, let κMn (s, p) be the cumulant generating function of a
multi-nomial random variable Mn with PMF p:
n
X
κMn (s; p) ≡ ln pi esai , (94)
i=1

1
Setting s = b
and multiplying by b:
  n n
1 X ai +b ln pi M
bκMn ;b = b ln e b (ai + bi ), (95)
b i=1
b
i=1

where bi ≡ b ln pi , i = 1, . . . , n. The RHS of (95) is the pseudo-dot product of the vector a and b, no matter
how b is defined. When each component of b is bi ≡ b ln pi , i = 1, . . . , n, then the RHS of (95) is the pseudo
expected value of Mn . This connection between the scaled cumulant generating function bκMn 1b ; b and


pseudo expected value in the log semi-field was first observed in McEneaney[8].
Recall the decomposition (11) of stoption premium mn (a) into the sum of a non-negative adaption
premium, E0Q [sN ∗ −1 − s0 ], and the independent exercise value, m̃∗n (a):

mn (a) = E0Q [sN ∗ −1 − s0 ] + m̃∗n (a), (96)

where for I an independent stopping time:


n
X n
X
m̃∗n (a) = p∗i E0Q [sI−1 − s0 + ai |I = i] = p∗i ai ≤ mn (a). (97)
i=1 i=1

In the LOPM, the lower suboptimal value m̃∗n (a) can be calculated from the larger optimized value mn (a).
Let p = 1/b > 0 be the reciprocal of the scale which measures the precision of the PDF of each increment
n
M
si − si−1 . The partial derivative of p 1
ai w.r.t. p is m̃∗n (a):
p
i=1
  !
n n n
∂  M ∂ X X
p 1
ai =
 ln epai = p∗i ai . (98)
∂p p ∂p i=1 i=1
i=1

19

Electronic copy available at: https://ssrn.com/abstract=3929583


n
M n
M
When all of the ai are equal to a, then for each n = 2, 3 . . ., a = a + b ln n, so a is affine in
b b
i=1 i=1
M n
both a and b. When this is not the case, then it is well-known that for each n = 2, 3 . . ., ai is strictly
b
i=1
convex in each ai , in the n−vector a, and in the base controller/bewilderment b > 0.
We have shown that the LOPM stoption premium is a pseudo-sum of n given floors ai :
n
M
mn (a, b) = ai (99)
b
i=1

Now suppose that each floor is itself a pseudo-sum of n given floors f (ai , aj )
n
M
ai = f (ai , aj ) (100)
b
j=1

Substituting (100) in (99) leads to a once iterated pseudo-sum:


n
M n
M
mfn (a, b) = f (ai , aj )
b b
i=1 j=1
n n
!
M X f (ai ,aj )
= b ln e b
b
i=1 j=1
n X
n
!
X f (ai ,aj )
= b ln e b . (101)
i=1 j=1

Thus a 2−fold pseudo-sum is related to a 2−fold sum via the appropriate insertion of exponentials and
logarithms. By extension, an n−fold pseudo-sum is related to an n−fold sum analogously:

n n n n X
n n
!
M M M X X f (ai ,ai ,...,ain )
1 2
mfn (a, b) = ... f (ai1 , ai2 , . . . , ain ) = b ln ... e b . (102)
b b b
i1 =1 i2 =1 in =1 i1 =1 i2 =1 in =1

5.3 Solution for the Critical Price Change Vector


Generalizing the conditional stoption value (61) from time n − 3 to time i:
n
M n
M
(n−i)
CVi = si − s0 + aj = si − si−1 + si−1 − s0 + aj , i = 1, . . . , n − 1. (103)
b b
j=i+1 j=i+1

(n−i)
For i = 1, . . . , n − 1, CVi denotes the stoption’s LOPM continuation value at time i, conditional on no
(0)
prior exercise. At time n, the continuation value is CVn = −∞, so that one is forced to exercise into the
reward sn−1 − s0 + an . The stoption’s exercise value at time i = 1, . . . , n is si−1 − s0 + ai , and it is also
conditional on no prior exercise. For i = 1, . . . , n − 1, we implicitly define the critical price change δi∗ as

20

Electronic copy available at: https://ssrn.com/abstract=3929583


(n−i)
the price change si − si−1 that equates the stoption’s LOPM continuation value CVi in (103) to the
stoption’s exercise value si−1 − s0 + ai :
n
M
δi∗ + si−1 − s0 + aj = si−1 − s0 + ai , i = 1, . . . , n − 1. (104)
b
j=i+1

For i = n, we explicitly define δn∗ = ∞. For i = 1, . . . , n − 1, we can explicitly solve (104) for δi∗ :
n
M
δi∗ = ai − aj . (105)
b
j=i+1

Thus, the LOPM generates the simple explicit formula (105) for each critical price change prior to maturity.
In words, if at some time i prior to maturity the price change si − si−1 happens to realize to the difference
Mn
between the current floor ai and the pseudo-sum of the future floors aj , then the stoption owner is
b
j=i+1
indifferent between exercising and continuing.
Of course, for a logistically distributed increment, the ex-ante probability of this particular realization
is zero. An optimal early exercise strategy at time i is to exercise if si − si−1 < δi∗ , else continue. At the
penultimate time n − 1, the critical price change simplifies to δn−1 = an−1 − an . Thus, if the last two floors
happen to be equal, then the stoption owner exercises on a drop over the penultimate period, otherwise he
or she continues. Suppose that at some time i prior to maturity, the current floor ai is not only higher than
Mn
any future floor, but that ai is also higher than their pseudo-sum aj . Then the critical price change
b
j=i+1
δi∗ is positive, so it is possible that the stoption owner surrenders a small enough gain si − si−1 > 0 for the
relatively high ai . A sufficient condition to never surrender a gain at time i = 1, . . . , n − 1 is that a future
floor is higher. If it so happens that all floors are equal, then the exercise-boundary for price changes (105)
simplifies to:
δi∗ = −b ln(n − 1 − i), i = 1, . . . , n − 2. (106)
Notice that this exercise boundary formula is independent of the common level of the floors and that it
runs from time 1 to time n − 2, not time n − 1. At time n − 1 and when the last two floors are equal,

(106) does not apply and δn−1 = 0, as mentioned earlier. Thus, any drop in the underlying’s price over the
penultimate period and under equal floors is exercised away. However, for equal floors, a prior drop may
not be exercised away if it is sufficiently small.

5.4 Semi-static Replication with Cash and One Vanilla Single Period Call
Setting i = 1 in (103) gives the stoption’s LOPM continuation value at time 1:
n
M
(n−1)
CV1 = s1 − s0 + aj . (107)
b
j=2

Thus, an instant before this first exercise opportunity, the stoption’s value is the larger of the exercise
value a1 and the stoption’s LOPM continuation value:
(n) (n−1)
CV1− = a1 ∨ M1 . (108)

21

Electronic copy available at: https://ssrn.com/abstract=3929583


Substituting (107) in (108), this value matches the payoff from a portfolio of a single-period bond and a
single-period vanilla call:
 
M n
(n)
CV1− = a1 ∨ s1 − s0 + aj 
b
j=2
   
n
M
= a1 + 0 ∨ s1 − s0 + a1 − aj 
b
j=2

= a1 + (0 ∨ (s1 − (s0 + δ1∗ ))), (109)


from (104) with i = 1. The single period bond has face value a1 ∈ R, while the single-period vanilla call
is struck δ1∗ dollars above the initial spot s0 . If the first price change is sufficiently low, i.e. s1 − s0 < δ1∗ ,
then the stoption is optimally exercised at time 1. In this case, the call in the replicating portfolio expires
worthless and the bond provides the stoption reward. If instead, s1 − s0 > δ1∗ , then the proceeds from
liquidating the bond call portfolio are just sufficient to buy a 1-period bond with face value a2 and a
1−period vanilla call struck δ2∗ dollars above s1 . Thus, this stoption replicating strategy is self-financing,
so long as the stoption is optimally kept alive. The stoption replicating strategy is dynamic in discrete-time
and semi-static in continuous-time.

6 Representation of the Stoption Premium via Convex Duality


Recall that the stoption’s initial premium must be convex in the floors. In this section, we use convex
duality to give an alternative representation of a stoption’s initial premium. This representation will be
used in the next section to generate several financial results of interest.
Let a ∈ Rn be the n−vector of floors and let p be the n−vector of risk-neutral probabilities of stopping
Pn Pn
at time i. In other words, each element pi of p is non-negative and pi = 1. Let H(p) ≡ − pi ln pi be
i=1 i=1
the Shannon entropy function. The function H(p) is positive and concave in the vector p of probabilities.
Let b > 0 be the bewilderment of each price change in the underlying security.
Consider the following maximization problem:
" n # " n n
# " n #
sup X sup X X sup X
p∈[0,1]n pi ai + bH(p) =p∈[0,1]n p i ai − b pi ln pi =p∈[0,1]n pi (ai − bi ) , (110)
i=1 i=1 i=1 i=1

where bi ≡ b ln pi . Since p is an n−vector of probabilities, the maximization over p ∈ [0, 1]n in (110) is
subject to the constraint:
Xn
pi = 1. (111)
i=1
n
M
Since pi > 0, we have bi ≡ b ln pi < 0 and it is easy to show that bi = 0:
b
i=1

n n
! n
! n
!
M X bi X b ln pi X
bi ≡ b ln e b = b ln e b = b ln pi = b ln (1) = 0. (112)
b
i=1 i=1 i=1 i=1

22

Electronic copy available at: https://ssrn.com/abstract=3929583


Let λ be a Lagrange multiplier and let L(p, λ; b) be the following Lagrangian:
n n
!
X X
L(p, λ; b) ≡ pi (ai − bi ) + λ 1 − pi . (113)
i=1 i=1

The Lagrangian is infinitely differentiable in each element of p, so taking the first derivative w.r.t. pi
(recalling that bi = b ln pi ) and setting it equal to zero implies that the optimizing p, denoted p∗ satisfies:
ai − b(ln p∗i + 1) − λ = 0. i = 1, . . . , n. (114)
Equivalently:
ai = b + λ + b∗i , i = 1, . . . , n, (115)
where b∗i ≡ b ln p∗i < 0 for i = 1, . . . , n. Pseudo-summing (115) over i and using the distributivity of + over
⊕:
n
M n
M
ai = (b + λ + b∗i )
b b
i=1 i=1
n
M
= b+λ+ b∗i
i=1
= b + λ, (116)
from (112) with pi = p∗i . Substituting (116) in (115) implies the following cross-period constraint on the
b∗i ≡ b ln p∗i :
n
M

ai − b i = ai i = 1, . . . , n. (117)
b
i=1
Notice that the RHS of (117) is independent of both b∗ and i. Solving (117) for the p∗i in the b∗i would
give the Gibbs PMF (80). However, we can determine the supremum in (110) without even determining
the p∗i , since substituting (117) in (110):
n
X n
X n
X n
M n
M n
X n
M
p∗i ai ∗
+ bH(p ) = p∗i (ai − b∗i ) = p∗i ai = ai p∗i = ai . (118)
b b b
i=1 i=1 i=1 i=1 i=1 i=1 i=1

Mn
Thus, we can represent the LOPM stoption premium ai as the solution to the constrained maximiza-
b
n  i=1
sup P Pn
tion p∈[0,1]n pi ai + bH(p; b) , subject to the constraint that pi = 1. This representation has financial
i=1 i=1
implications, as we will see in the next section.

7 Financial Representations of LOPM Stoption Premium


7.1 Adaption Premium
The optimized objective (118) is a two term decomposition of the LOPM stoption premium:
n
M n
X
ai = p∗i ai + bH(p∗ ). (119)
b
i=1 i=1

23

Electronic copy available at: https://ssrn.com/abstract=3929583


The first term on the RHS of (119) is the value m̃∗n of the mismanaged stoption where the random stopping
time I used to exercise it is independent of the underlying security price path. Thus, the first term on
the RHS of (119) is the expected value of the payoff sI−1 − s0 + aI∗ , while the whole RHS of (119) is the
expected value of the payoff sN ∗ −1 − s0 + aN ∗ . with both expectations calculated using Gibb’s PMF (80).
Thus, the stopping times N ∗ and I are identically distributed random variables, but they radically differ
in terms of their dependence on the underlying security price path. The last term on the RHS of (119)
is the adaption premium, i.e. the extra value due to this differing dependence. Since Shannon entropy
is positive, the stoption has greater value than the corresponding claim with independent stopping times.
Moreover, since Shannon entropy H(p∗ ) is maximized when the Gibbs probabilities p∗i are equal, moving
from a stoption with constant floors a to a stoption with varying floors but with the same average floor
n
p∗i ai = a, will lower the LOPM premium.
P
i=1

7.2 Vega of the LOPM Stoption Premium


From (12), (85), and (119), the adaption premium has three representations:
n
∂ M
−E0Q [sN ∗ − sN ∗ −1 ] = b ai = bH(p∗ ). (120)
∂b b
i=1

The last equality in (120) implies that in the LOPM, the stoption’s vega is just the Shannon entropy of
the Gibb’s probabilities:
n
∂ M
ai = H(p∗ ). (121)
∂b b
i=1
Suppose that a floor vector a must average to some given mean a ∈ R using the Gibb’s PMF p, which
depends on a. Requiring a · p = a, (121) implies that the stoption’s vega in the LOPM is maximized by
setting each floor equal to a.

7.3 Generalized Put-Call-Parity in the Logistic Model


Suppose that we change the indexing variable on the RHS of (117) from i to j:
n
M
ai + (−b∗i ) = aj , i = 1, . . . , n, (122)
b
j=1

where b∗i ≡ b ln p∗i . Re-arranging (122):


n
M n
M
−b∗i = −ai + aj = (aj − ai ), i = 1, . . . , n, (123)
b b
j=1 j=1

by the distributivity of + over ⊕. Thus, −b∗i > 0 is the LOPM value of a stoption with floors aj − ai
for j = 1, . . . n. Since one of the floors vanishes, we think of −b∗i > 0 as the value of a long position in a
generalized option. Substituting (123) in (122) implies:
n
M n
M
ai + (aj − ai ) = aj , i = 1, . . . , n. (124)
b b
j=1 j=1

24

Electronic copy available at: https://ssrn.com/abstract=3929583


Equation (124) is a direct consequence of the distributivity of + over ⊕b , i.e. for any λ0 ∈ R:
n
M n
M
λ0 + (aj − λ0 ) = aj , λ0 ∈ R. (125)
b b
j=1 j=1

Equation (124) merely illustrates (125) n times with λ0 = a1 , λ0 = a2 , . . . , λ0 = an .


The LHS of (124) is the initial value of a portfolio consisting of a riskless security initially worth ai ∈ R
Mn
and a generalized option initially worth (aj − ai ) in the LOPM. The generalized option is itself an
b
j=1
n−period stoption, but the i−th generalized option is required to have its i−th floor vanish. The initial
premium of each generalized option is a function of n − 1 generalized moneyness levels δ1 , . . . , δn−1 . When
n = 2, there are two options, each a function of one moneyness level. If the call moneyness is δ, then the
put moneyness is −δ. More generally for any finite n = 2, 3 . . ., we will show that the grand sum over all
n generalized options of the n × (n − 1) non-zero floors, each a function of n − 1 moneyness levels, must
vanish. For example, if n = 3, the grand sum of the 6 non-zero floors, each a function of moneyness levels
δ1 and δ2 , must vanish. We will also show that the grand pseudo-sum of the generalized option premia
must vanish as well.
Recall that the LHS of (124) is the initial value of a portfolio of a bond and a generalized option. The
distributivity of + over ∨ causes the payoff from the portfolio to match the payoff from the stoption on
the RHS of (124). It follows from no arbitrage that the premia must also align, which is (124). We can
think of (124) as an n−period generalization of put-call-parity (PCP). It reduces to the standard PCP by
setting n = 2 in (124).
In (123), we interpreted −b∗i as the LOPM initial premium of an n−period stoption. In fact, both
terms on the LHS of (124) can be interpreted as the LOPM initial premium of an n−period stoption. A
bond paying ai dollars at time i has the same payoff as an n−period stoption where the i−th floor is ai
and the n − 1 other floors all equal −∞:
n
M aj
ai = b ln(1j=i e b ), i = 1, . . . , n. (126)
b
j=1

The payoffs of the n generalized options in the generalized PCP (124)


Pn have a relationship to each other.
For each i, the ordinary sum of each generalized option’s floorsPn Pn j=1 (aj − ai ). Summing across i, the
is
grand sum of all of the floors across all generalized options is i=1 j=1 (aj −ai ). This grand sum vanishes,
since distributing the j sum across the minus sign implies:
n X n n n
! n n
X X X X X
(aj − ai ) = aj − nai = n aj − n ai = 0. (127)
i=1 j=1 i=1 j=1 j=1 a=1

Notice that dividing the term after the first equality by −n2 implies:
n n
! n n
!
1 X X 1X 1X
− 2 aj − nai = ai − aj = 0. (128)
n i=1 j=1 n i=1 n j=1

In words, the simple average of the deviations from the simple average vanishes. This well-known result is
what leads to the grand sum vanishing in (127).

25

Electronic copy available at: https://ssrn.com/abstract=3929583


8 Insured Futures Contract
A futures contract has zero premium at inception and afterwards. Long and short positions in futures are
both subject to daily marking-to-market. Each day, the long position pays the short position the change
in the futures price. These changes can be positive or negative, and the daily marking-to-market payments
continue until the futures contract is either closed or matures.
A stoption is a hybrid between a futures contract and a Bermudan put option. We have thus far treated
the stoption’s floors as given and we have computed the resulting LOPM stoption premium in closed-form.
In this section, we instead force the initial stoption premium to be zero in order to promote the similarity to
futures contracts. We also demand that all of the n floors of the stoption be non-positive. When a set of n
non-positive floors leads to zero initial stoption premium, we call the floors feasible and we call the stoption
an insured futures contract. A simple way to create a set of n feasible floors is to choose a probability
mass function (PMF) of a categorical random variable with n outcomes. For i = 1, . . . , n, let pi ≥ 0 denote
Pn
the probabilities in this PMF so that pi = 1. To determine a set of feasible floors, one must also know
i=1
the scale parameter b > 0 of the zero mean logistic distribution governing the n independent daily price
changes in the stoption’s underlying security. One then forces the floors ai to be calculated as:

ai = b ln pi ≡ bi ≤ 0, i = 1, . . . , n. (129)

It follows from (112) and (129) that:


n
M
ai = 0. (130)
b
i=1

In words, the resulting LOPM stoption premium vanishes.


The intuition behind this result arises from the decomposition (119) of the stoption premium into
n
P
the ordinary sum of intrinsic value pi ai and time value bH(p) > 0. Since all of the n floors ai are
i=1
non-positive, the stoption holder is being forced to stop into a non-positive floor. As a result, the insured
n
P
futures contract’s intrinsic value pi ai is also non-positive. However, the positive time value bH(p) > 0
i=1
of the insured futures contract raises its premium to zero. This approach of specifying a PMF is not the
only way to find a set of feasible floors, but it has the advantage that each pi in the PMF chosen is the
risk-neutral probability that the optimal stopping time will be i. For the reminder of the paper, we return
to the setting where the n stoption floors are n given real numbers, so that the stoption premium need not
vanish.

9 Analogy with the Positive Gain Beta CAPM


In this section, we first develop the gain CAPM, which is very similar to the more standard return CAPM.
We then restrict the gain CAPM so that all gain betas are positive. We show that the first order conditions
arising in the positive gain beta CAPM are isomorphic to the first order conditions arising when valuing
a stoption via convex duality.
Consider a single period economy with n + 1 assets, where n is a finite natural number. The zero-eth
asset is riskless. One dollar invested in the riskless asset at the beginning of the period becomes the

26

Electronic copy available at: https://ssrn.com/abstract=3929583


deterministic amount R0 > 0 at the end of the period. When the riskless interest rate is positive, then the
riskless gross return R0 > 1, while R0 ∈ (0, 1] when the riskless interest rate is not positive.
The remaining n assets are each individually risky and are each in positive supply at both the beginning
and the end of the period. Recall that the number of shares of each risky asset is one. For i = 1, . . . , n,
let Vi0 > 0 be the known initial market capitalization of the i-th risky asset and let ViT ≥ 0 be its random
terminal market capitalization.
Let VM 0 be the sum of the known initial market capitalizations of the n risky assets and let VM T be
the sum of the n random terminal market capitalizations:
n
X n
X
VM 0 = Vi0 > 0 VM T = ViT ≥ 0. (131)
i=1 i=1

Following standard convention, we refer to a combined holding in all of the risky assets according to their
market capitalizations as a position in the market portfolio. In our setting, the market portfolio is a holding
of one share of each risky asset.
Let P be a probability measure. We assume that the random future values of all of the n risky assets
each have finite P mean. It follows from taking expected values in (131) that the random future value of
the market portfolio also has a finite P mean. We also assume that for any pair of future risky asset values
ViT and VjT , the mean of the product is finite. It follows that all covariances of future values are finite and
that all variances of future values are finite. As a result, all covariances of future values with the value of
the market portfolio are finite. Moreover, the variance of the future value of the market portfolio must be
finite.
The excess gain on asset i is defined as ViT − Vi0 R0 for i = 1, . . . , n. This random excess gain is achieved
at zero cost by borrowing the positive cost of each asset. The expected excess gain on asset i is:

Ei ≡ E0P ViT − Vi0 R0 , i = 1, . . . , n. (132)

The random excess gain on the market portfolio M is defined as:


n
X n
X n
X
VM T − VM 0 R0 = ViT − R0 Vi0 = (ViT − R0 Vi0 ), (133)
i=1 i=1 i=1

from (131) and algebra. Thus, the expected excess gain on the market portfolio M is:
n
X n
X
E0P (VM T − VM 0 R0 ) = E0P (ViT − R0 Vi0 ) = Ei . (134)
i=1 i=1

The variance of the excess gain on the market portfolio M is:

VarP (VM T − VM 0 R0 ) = VarP (VM T ). (135)

We treat the expected excess gain on asset i and the covariance of excess gains between assets i and j as
directly observed at time 0. The expected excess gain on asset i is measured in dollars, while the covariance
of excess gains between assets i and j is measured in dollars squared.
Let hi ∈ R be the holding of asset i, for i = 1, . . . , n. When the position in each asset is fully
financed by borrowing or lending, then the random excess gain from holding hi units of each asset is

27

Electronic copy available at: https://ssrn.com/abstract=3929583


n
P
hi × (ViT − Vi0 R0 ). The expected excess gain from holding hi units of a leveraged zero-cost position in
i=1
asset i is hi × Ei . Let h be the vector of asset holdings. The expected excess gain associated with asset
Pn
holdings vector h is hi × Ei . This expected excess gain is achieved at zero cost. The variance of excess
i=1
gains associated with asset holdings vector h is:
n
! n X
n
X X
VarP0 hi × (ViT − Vi0 R0 ) = hi hj CovP0 (ViT , VjT ). (136)
i=1 i=1 j=1

The market portfolio is defined as the holdings vector h = 1. As a result, the variance of excess gains
associated with the market portfolio arises by setting vector h = 1 in (136):
n
! n
! n X n
X X X
VarP0 (ViT − Vi0 R0 ) = VarP0 ViT = VarP0 (VM T ) = CovP0 (ViT , VjT ). (137)
i=1 i=1 i=1 j=1

We assume that the risk-averse investor maximizes expected excess gain of their holdings by choice
of h, subject to the constraint that the variance of excess gains from these holdings be some acceptably
n
small level σ 2 > 0. The investor’s problem is to maximize the expected excess gain
P
hi × Ei by choice of
i=1
portfolio holdings h ∈ Rn , subject to the following constraint:
n X
X n
hi hj CovP0 (ViT , VjT ) = σ 2 , (138)
i=1 j=1

where σ 2 > 0. Let Λ > 0 be a Lagrange multiplier and let L(h, Λ) be the Lagrangian for this constrained
minimization problem:
n
" n Xn
#
X X
L(h, Λ; σ 2 ) ≡ hi × Ei + Λ σ 2 − hi hj CovP0 (ViT , VjT ) . (139)
i=1 i=1 j=1

The investor must maximize L(h, Λ) by choice of the asset holdings vector h ∈ Rn and the Lagrange
multiplier Λ > 0. For i = 1, . . . , n, the first partial derivatives of the Lagrangian L(h, Λ; g) w.r.t. hi are:
n
∂ 2
X
L(h, Λ; σ ) = Ei − Λ hj CovP0 (ViT , VjT ). (140)
∂hi j=1

A necessary condition for optimality is that the gradient vanishes, ∂h∂ i L(h∗ , Λ; σ 2 ) = 0∀i, so that for the
optimal holdings h∗ :
Xn
Ei = Λ h∗j CovP0 (ViT , VjT ), i = 1, . . . , n. (141)
j=1

Notice that this first-order condition (141) is affine in the vector h∗ .


At the optimum, the market must be held. Setting h∗ = 1 in (141) simplifies it to:
n
X
Ei = Λ CovP0 (ViT , VjT ) i = 1, . . . , n. (142)
j=1

28

Electronic copy available at: https://ssrn.com/abstract=3929583


Covariance is a bi-linear operator so we can distribute the sum:
n
!
X
P
Ei = ΛCov0 ViT , VjT , i = 1, . . . , n. (143)
j=1

Substituting in (131) simplifies (143) further:

Ei = ΛCovP0 (ViT , VM T ) , i = 1, . . . , n. (144)

Suppose we divide both sides of (144) by the variance VarP0 (VM T ) of the market portfolio. Letting:

Ei
Ai ≡ , i = 1, . . . , n, (145)
VarP0 (VM T )

and:
CovP0 (ViT , VM T )
Bi ≡ , i = 1, . . . , n, (146)
VarP0 (VM T )
(144) becomes:
Ai = ΛBi i = 1, . . . , n. (147)
From the definition (146), gain betas sum to one:
 n

P
n n CovP0 ViT , VM T
X X CovP (ViT , VM T )
0 i=1
Bi = = = 1, (148)
i=1 i=1
VarP0 (VM T ) VarP0 (VM T )

from (131). We can use this result to identify Λ in (147). Summing (147) over i implies:
n
X n
X
Ai = Λ Bi = Λ, i = 1, . . . , n, (149)
i=1 i=1

from (148). Dividing (147) by Bi > 0 implies:


n
Ai X
=Λ= Ai i = 1, . . . , n, (150)
Bi i=1

from (149). Multiplying both sides of (150) by the variance VarP0 (VM T ) of the market portfolio and using
(145):
n
Ei X
= Ei i = 1, . . . , n, (151)
Bi i=1

In words, the ratio of reward Ei to risk Bi is the same across all assets i. The common ratio is the mean
Pn
excess gain Ei on the market portfolio. The RHS of (151) is implicitly the ratio of reward to risk for
i=1
the market portfolio since its risk relative to itself is one.

29

Electronic copy available at: https://ssrn.com/abstract=3929583


If we set σ 2 in (139) to the market portfolio variance VarP0 (VM T ), then the maximized mean gain is
n
Ei , since the optimal holding is h∗ = 1, i.e.:
P
that of the market portfolio,
i=1

n
X
L(1, Λ; VarP0 (VM T )) = Ei . (152)
i=1

We henceforth assume that the gain betas are all positive, i.e. Bi > 0, i = 1, . . . , n. Since the gain
betas sum to one from (148), we must have that each Bi is in the unit interval (0, 1). In words, the gain
betas Bi , i = 1, . . . , n in the positive gain beta CAPM are probabilities. Since investors are assumed to be
Pn
risk-averse, the mean excess gain of the market portfolio is positive, i.e. Ei > 0. In the positive gain
i=1
beta CAPM, we must also have from (151) that each mean gain is positive, Ei > 0, i = 1, . . . , n. As a
result, the definition (145) of Ai implies that each of the Ai are positive.
Now recall the FOC (115) arising in the convex dual optimization problem whose solution is the stoption
premium: If we set λ̃ = b + λ this FOC becomes:

ai = λ̃ + b∗i , i = 1, . . . , n. (153)

It is natural to compare the FOC (153) to the FOC (147) repeated here:

Ai = ΛBi i = 1, . . . , n. (154)

The Bi in (154) are between 0 and 1 and sum to one. The b∗i in (153) are the counterparts of Bi under
the map b ln g for g > 0, b > 0. Analogously, the b∗i in (153) are negative and pseudo-sum to zero. The
Ai in (154) are positive under our assumption that all Bi > 0. The stoption floors ai in (153) are the
counterparts of Ai under the map b ln g for g > 0, b > 0. Analogously, the stoption floors ai in (153) are
real-valued. From (149), the Lagrange multiplier Λ in (154) gets identified as the sum of the Ai . The
arithmetic used in this identification is conducted in the semi-field ([0, ∞); +, 0; ×, 1). The analog of this
semi-field under the generator b ln g for g > 0, b > 0 is the semi-field ([−∞, ∞); ⊕b , −∞; +, 0). In this
latter semi-field, pseudo-summing (153) over i and using the distributivity of + over ⊕:
n
M n
M
ai = (λ̃ + b∗i )
b b
i=1 i=1
n
M
= λ̃ + b∗i
i=1

= λ̃, (155)

since ni=1 b∗i = 0. Thus, the shifted Lagrange multiplier λ̃ in (153) gets identified as the pseudo-sum of
L
the ai . Recall from (152) that the maximized Lagrangian in the positive gain beta CAPM is the sum of
Pn
the mean excess gains, Ei > 0. Analogously from (118), the maximized Lagrangian in the convex dual
i=1
n
M
optimization problem (113) is the pseudo-sum of the floors, ai .
b
i=1

30

Electronic copy available at: https://ssrn.com/abstract=3929583


We conclude that the relationship at the optimum between each individual security mean excess gains
n
P
Ei > 0 and the market portfolio’s mean excess gains Ei > 0. in the linear positive gain beta CAPM is
i=1
n
M
isomorphic to the relationship at the optimum between the floors ai ∈ R and the stoption premium ai
b
i=1
in the pseudo-linear LOPM.

10 Application to Debit Valuation Adjustment (DVA)


Debit Valuation Adjustment (DVA) is meant to monetize the benefit that one counterparty has from the
ability to default on the other counterparty. In this section, we suppose that only one counterparty can
default and we present a situation where the benefit from this ability to default is literally the payoff from
a stoption with zero floors. Suppose that counterparty A is to receive the value in cash of a reference
security in n periods from counterparty B. The security has a real-value. We assume that counterparty
A can default but counterparty B cannot. If the security’s realization is negative, then A must pay B the
absolute value of the security. Since A can default, A may choose to not pay B anything in this case. This
is a well known example of A benefitting by defaulting.
Counterparty B is well aware ex ante that A may default on their obligations. As a result, suppose
that at the beginning of the trade, B is able to force A to initially agree to post cash collateral every time
in the future that the price of the security drops. The size of the collateral posting is to be the size of the
price drop. We say that B has imposed full collateralization on A. Of course the security price could rise
over a future period instead of dropping. In this case, the full collateralization agreement stipulates that A
receives a positive amount equal to the rise. In the absence of any default by A, this full collateralization
agreement causes A to receive the price change when it is positive and pay the absolute value of the price
change when it is negative. However even in the presence of a full collateralization agreement, A can still
default. In particular, A may choose to not pay the absolute value of a large enough drop. A default by
A causes the collateralization posting to stop and B will no longer be required to deliver the security to
A at expiration. For a small enough price drop, A won’t default in order to keep the option of defaulting
later open. In short, A’s ability to default in the presence of a full collateralization agreement gives A the
benefits conferred by a stoption with all floors zero. The initial DVA under n IID logistically distributed
price changes is:
M n
DV A0 (n) = mn (0, b) = 0 = b ln n. (156)
b
i=1

Recall that the exercise-boundary for price changes in this case is δi∗ = −b ln(n − 1 − i), i = 1, ldots, n − 1.
One can introduce exogenous bankruptcy costs for counterparty A, which potentially vary with the
time of A’s default. Exogenous bankruptcy costs suffered by A upon their default are captured by setting
the floors negative in our setting. In this case, the initial DVA is lowered and just becomes the pseudo-sum
of the n negative floors:
Mn
DV A0 (n) = mn (a, b) = ai , a ≤ 0. (157)
b
i=1

If for any reason, there is a bankruptcy benefit rather than cost at any time i, then (157) still holds, but
with the restriction on a removed.

31

Electronic copy available at: https://ssrn.com/abstract=3929583


11 Replacing Max with Min and Helmholtz Free-Energy
Recall from (6) that the premium arising from replacing max with min in the payoff is related to the
stoption premium. In this section, we first examine the premium arising from replacing max with min
when the underlying’s spot price increments are IID logistic with mean zero and scale b > 0, i.e. when:
N
X −1 N
X −1
sN −1 − s0 = (si − si−1 ) = bZi . (158)
i=1 i=1

We then show that this new pricing formula is identical to Helmholtz free-energy.

11.1 Premium when Replacing Max with Min


Substituting (158) in (5) implies that the premium arising when max is replaced by min has the probabilistic
representation:
"N −1 #
min min X
m− Q Q
n (a, b) ≡N ∈[1,2,...,n] E0 [sN −1 − s0 + aN ] =N ∈[1,2,...,n] E0 bZi + aN . (159)
i=1

This premium can be related to the stoption premium:


( "N −1 #)
min X
m− Q
n (a, b) ≡ N ∈[1,2,...,n] E0 (−1)(−1) bZi + aN
i=1
"N −1 #
max Q
X
= − N ∈[1,2,...,n] E0 (−bZi ) − aN
i=1
"N −1 #
max X
= − N ∈[1,2,...,n] E0Q bZi − aN , (160)
i=1

since each Zi is symmetric. Recall from (37) and (62) that the n−period stoption premium is a log sum
exponential:
"N −1 #
max X  a1 a2 an

Q
N ∈[1,2,...,n] E0 bZi + aN = b ln e + e + . . . + e
b b b , ai ∈ R, i = 1, . . . , n, b > 0. (161)
i=1

Using (161) in (160):

min
m−
n (a, b) ≡ E Q [s
N ∈[1,2,...,n]− s 0 + aN ]
 a10 N −1 
− b − abn
= −b ln e + ... + e
= −[(−a1 ) ⊕b . . . ⊕b (−an )], (162)

from the definition (45) of ⊕b . For E1 , E2 ∈ (−∞, ∞], define a new binary operation ⊕−b by:
 E1 E2

E1 ⊕−b E2 ≡ −b ln e− b + e− b = −(−E1 ⊕b −E2 ), b > 0. (163)

32

Electronic copy available at: https://ssrn.com/abstract=3929583


This new binary operation has ∞ as its neutral element and is associative. Substituting (163) in (162):

min
m− Q
n (a, b) ≡ N ∈[1,2,...,n] E0 [sN −1 − s0 + aN ]
= a1 ⊕−b . . . ⊕−b an
M n
≡ ai , a ∈ Rn . (164)
−b
i=1

Thus, the premium resulting from replacing max with min in the payoff can be captured as the usual
pseudo-sum, but replacing b > 0 with −b < 0. The pseudo-summands ai should be referred to as caps, not
floors.

11.2 Helmholtz Free-Energy


In this section, we consider what appears to be a completely unrelated problem. Let S > 0 denote entropy.
Let U (S) be internal energy where U : R+ 7→ R+ is increasing and convex in S. Let T > 0 denote
temperature, which will be held constant. We consider a system in contact with a heat bath. A standard
argument in statistical mechanics is that temperature T is much easier to measure than entropy S. As a
result, Helmholtz[11] considered the following optimization problem:
inf
F (T ) ≡ S [U (S) − T S]. (165)

The function of S and T being minimized over on the RHS of (165) is called free-energy, while the function
of T on the LHS of (165) is called Helmholtz free-energy. The Helmholtz free-energy describes the useful
work obtainable from a closed thermodynamic system at a constant temperature and volume.
Let e ∈ Rn be an n−vector of energies and let P be an n−vector of probabilities. In other words, each
Pn
element Pi of P is non-negative and Pi = 1. Let:
i=1

n
X
S(P) ≡ −kB Pi ln Pi (166)
i=1

be the entropy function, as it appears in statistical mechanics, as opposed to information theory. The
Boltzmann constant kB in (166) is positive. The function S(p) is positive and concave in the vector P of
probabilities. With S depending on the probability vector P, we set U (S(P)) equal to the mean energy
Pn
Pi ei :
i=1
n
X
U (S(P)) = Pi e i . (167)
i=1

With S and U (S) respectively given as functions (166) and (167) of the probability vector P, we con-
sider minimizing free-energy U (S) − T S over probabilities Pi > 0 instead of over entropy S > 0. The
minimization is subject to the constraint:
Xn
Pi = 1. (168)
i=1

33

Electronic copy available at: https://ssrn.com/abstract=3929583


Thus, we consider the following constrained minimization problem:
" n n
#
inf X X
F̃ (E; b) ≡P∈[0,1]n Pi e i + b Pi ln Pi , (169)
i=1 i=1

where b ≡ kB T is a positive constant and the Pi solve (168).


Let λ be a Lagrange multiplier and let L(P, λ; b) be the following Lagrangian:
n n n
!
X X X
L(P, λ; b) ≡ Pi e i + b Pi ln Pi + λ 1 − Pi . (170)
i=1 i=1 i=1

The Lagrangian is infinitely differentiable in each element of P, so taking the first derivative w.r.t. pi and
setting it equal to zero implies:

Ei + b(ln Pi∗ + 1) − λ = 0, i = 1, . . . , n, (171)

or equivalently:
ei = −b ln Pi∗ + λ − b, i = 1, . . . , n. (172)
Pseudo-summing (172) over i:
n
M n
M
ei = (−b ln Pi∗ + λ − b)
−b −b
i=1 i=1
n
M
= λ−b+ (−b ln Pi∗ ), (173)
−b
i=1

from the distributivity of + over ⊕−b . However, from the definition (164) of a repeated pseudo-sum:
n n
! n
!
M X −b ln Pi∗ X
(−b ln Pi∗ ) = −b ln e− b = −b ln Pi∗ = −b ln 1 = 0, (174)
−b
i=1 i=1 i=1

from requiring the constraint (168) to hold for the optimized probabilities. Substituting (174) in (173)
implies that:
Mn
λ−b= ei . (175)
−b
i=1

Substituting (175) in the FOC (172) and re-arranging:


n
M
ei + b ln Pi∗ = ei , i = 1, . . . , n. (176)
−b
i=1

Now the minimized objective function (110) is:


n
X n
X n
M n
M
F̃ (e; b) = p∗i (ei + b ln pi ) = p∗i ei = ei , e ∈ Rn , b > 0, (177)
−b −b
i=1 i=1 i=1 i=1

34

Electronic copy available at: https://ssrn.com/abstract=3929583


n n
M
Pi∗
P
since = 1. Substituting the definition (163) of ei into (177) implies that the Helmholtz
i=1 −b
i=1
free-energy is: !
n
X
F̃ (e; b) = −b ln e−ei /b , e ∈ Rn , b > 0. (178)
i=1

It follows from (158) and (162) that Helmholtz free-energy has the following probabilistic representation:
N −1
min X
F̃ (e; b) =N ∈[1,2,...,n] E0Q [bZi + eN ] , E ∈ Rn , b > 0, (179)
i=1

where the Zi are IID standard logistic.


Notice that by using the relevant pseudo-arithmetic, we were able to determine the minimized free-
energy F̃ (e; b) as a log-sum-exponential or pseudo-sum without first determining the optimized probabilities
Pi∗ . To determine these probabilities, first re-arrange (176):
n
M
−b ln Pi∗ = ei − ei , i = 1, . . . , n. (180)
−b
i=1

With energies real, this is just the pseudo-probability normalization condition in the commutative monoid
((−∞, ∞]; ⊕−b , ∞). Solving (180) for Pi∗ :
n
M
ei − ei
−b ei
i=1 e− b
Pi∗ =e − b = P
n e
, i = 1, . . . , n, (181)
− bi
e
i=1

from (163). The fraction on the RHS of (181) is called the Gibbs probability mass function (PMF) and
its denominator is called the partition function. The n positive quantities on the LHS of (180) can be
called pseudo-Gibbs probabilities, where the generator of the pseudo-arithmetic is G(P ) = −b ln P . These
pseudo-Gibbs probabilities pseudo-sum to zero because the real-valued energy levels ei act as caps on the
logistically distributed increments bZi , rather than floors.

12 Nonzero Constant Riskfree Interest Rate


Thus far, we have assumed that the riskfree interest rate is zero. In this section, we assume that the
riskfree interest rate is a non-zero constant instead. We assume that the underlying is a futures price so
that it is still a Q martingale. However, the stoption payoff changes so that past gains earn interest. Let
R > 0 be the gross return per period on the riskless asset. One dollar invested at the beginning of a period
becomes R dollars at the end of the period. The probabilistic representation (1) for the spot value of a
stoption becomes:
"N −1 #
max X
msn (a) =N ∈[1,2,...,n] E0Q R−N (si − si−1 )RN −i + aN , n = 1, 2 . . . . (182)
i=1

35

Electronic copy available at: https://ssrn.com/abstract=3929583


Recall that our logistic model assumed under zero interest rates that price changes are IID logistic with
zero mean and scale b > 0. Under non-zero interest rates, we retain independence and that increments
are logistically distributed with zero mean, but we now allow different increments to have different scales.
In particular, the scale parameter of increment si − si−1 becomes bRi > 0. Hence, if interest rates are
positive, the further an increment is in the future, the more volatile it is. Still letting Zi be IID standard
logistic random variables, we now have that each increment si − si−1 is represented as:

si − si−1 = bZi Ri , i = 1, . . . , n. (183)

Substituting (183) in (182) implies:


"N −1 #
max X
msn (a) = N ∈[1,2,...,n] E0Q R−N (bZi Ri )RN −i + aN
i=1
"N −1 #
max X
= N ∈[1,2,...,n]
Q
E0 bZi + aN R−N , n = 1, 2 . . . . (184)
i=1

It follows that the only change to the closed-form pricing formula for the stoption’s initial premium is that
the floors must be present-valued:
M n
s
ai R−i .

mn (a) = (185)
b
i=1

It follows that under positive interest rates, the stoption premium is lower than under zero interest rates.
The stoption premium is decreasing in R > 0 and still increasing in b > 0. As a result, the usual intuition
regarding the tradeoff between the time value of money and the volatility value of an option is available if
a structured product payoff embeds a stoption payoff, rather than a European option payoff.

13 Summary and Future Research


We introduced a new derivative security called a stoption. We gave a probabilistic representation for its
premium as the solution to an optimal stopping problem. We initially presented general properties of
the stoption premium. We then assumed that the stoption’s underlying security price followed a discrete-
time random walk. When the underlying’s price increments are IID, the initial premium of an n-period
stoption is the result of iterating a single-period married put’s value function on itself n − 1 times. We then
introduced a logistic model for the underlying security’s price changes. This LOPM allowed us to value
the stoption premium in closed-form. We also determine the optimal stopping strategy and a replication
strategy using one period options. Convex duality was used to decompose the LOPM stoption premium,
to determine vega, and to develop a generalized put call parity. Finally, we presented an application to
DVA (debit valuation adjustment) under full collateralization.
The LOPM is a discrete-time model, so it is natural to wonder if there is a supporting continuous-time
model. Carr and Torricelli[6] give two supporting continuous-time martingale models which both have
logistic marginals at every continuous forward time. The first is a pure-jump additive process, while the
second is a driftless diffusion. Recall that the stoption payoff could be replicated by rolling over one-period
vanilla calls. If these vanilla calls do not trade outright, then one can adopt the supporting driftless
diffusion model and replicate each vanilla call via dynamic trading in the stoption’s underlying security

36

Electronic copy available at: https://ssrn.com/abstract=3929583


and a riskless asset. It follows that the stoption’s payoff can also be replicated in this manner. Future
research can investigate the details of this replicating strategy and whether there are other supporting
continuous-time models.
In this paper, a logistic model was used to create the log-plus semi-field that was conveniently used
to value an n−period stoption. It is natural to wonder whether other distributions might lead to other
generators besides log that can alternatively be used to value a stoption. Here, the answer is a qualified no.
Carr and Costa[5] show that if a commutative monoid over [−∞, ∞) is required to be consistent with a
probability distribution governing the underlying security price at the end of a single period, then the only
distribution available is the logistic one used in this paper. However, they also show that if the support set
for the end of period price is not the real line, then other distributions are consistent with a commutative
monoid. In particular, if the support set for the end of period price is [0, ∞), then a 2 parameter continuous
distribution called conjugate power Dagum is consistent with a commutative monoid. Future research can
examine other support sets for final price and also consider non-commutative monoids.
Future research can examine the limit of the LOPM pricing formula as the time between observations
tends to zero. One thus transitions the stoption from Bermudan exercise to American exercise. If this
limit exists, we anticipate that the stoption premium is a pseudo-integral instead of a pseudo-sum.
The logistic model caused the married put value function to be a pseudo-sum. Pseudo-sums are just
one example of a function which has a closed-form for its iteration. Future research can attempt to find
other distributions with real support whose corresponding married put value function iterates in closed
form.
The analogy between the LOPM stoption valuation via convex duality and the positive gain beta CAPM
arose at the level of first-order conditions in a constrained optimization. In any optimization with a single
constraint, a necessary condition is that the gradient of the objective be proportional to the gradient of the
constraint at the optimum. Remarkably, the constrained optimization that leads to stoption premium via
convex duality recasts this necessary condition in a different arithmetic. At the optimum, the gradient of
the objective is a translation of the gradient of the constraint, instead of a multiple. Future research should
attempt to construct a pseudo-Lagrangian. Pseudo-differentiating this pseudo-Lagrangian and setting the
result to negative infinity should lead to first order conditions with pseudo gradients being translations,
rather than multiples.
One can also attempt to expand the number of underlying securities above one. For example if we now
treat the underlying security price as the value of spread or difference of two orthogonal security values,
then assuming the same Gumbel distribution for each of the 2 securities is consistent with assuming a
logistic distribution for the difference. In the interests of brevity, these extensions to stoption valuation
are best left for future research.

Appendix: Stoption Premium is a Generating Function


Recall the definition (186) of a pseudo-generating function: ĉ(z, b):

M
ĉ(z, b) ≡ (cj + zj), z ∈ R−∞ , b > 0. (186)
b
j=0

In this appendix, we show how to invert the pseudo-generating function ĉ(z, b) for a coefficient cj . We also
show that the pseudo-generating function of the pseudo-convolution of two sequences is just the ordinary
sum of the two pseudo-generating functions.

37

Electronic copy available at: https://ssrn.com/abstract=3929583


The inversion is most easily accomplished using non-Newtonian differentiation ( see Grossman and
·
Katz,[7]), but we will just use ordinary differentiation. Applying e b to (78):
∞ ∞
ĉ(z,b) X cj +zj X cj z j
e b = e b = e b eb . (187)
j=0 j=0

z
To recover cn , suppose that we now differentiate (187) w.r.t. e b n times:
 n ∞
d ĉ(z,b) X cj z j−n
z e b = n!e b e b . (188)
de b j=n

The LHS is the n−th non-Newtonian partial derivative of ĉ(z, b) w.r.t. z ∈ R−∞ . If we now set z = −∞
and divide by n!, we get:  n
1 d ĉ(z,b) cn
z e b =eb, (189)
n! de b z=−∞

since 00 = 1. Thus finally, for each n = 0, 1, 2 . . ., the pseudo-generating function ĉ(z, b) is inverted by:
  n 
1 d ĉ(z,b)
cn = b ln z e b . (190)
n! de b z=−∞

We now suppose that each element cj of the sequence {cj , j = 0, 1, 2, . . .} is a pseudo-convolution of


infinite sequences {dj , j = 0, 1, 2, . . .} and {ej , j = 0, 1, 2, . . .}, i.e.:
j
M
cj = (di + ej−i ), j = 0, 12, . . . . (191)
b
i=0

When (191) is substituted in (186), we call ĉ(z, b) the generating function of the pseudo-convolution. We
now show that the generating function of the pseudo-convolution of two sequences is just the ordinary sum
of the two generating functions. Substituting (191) in (186) and then using the distributivity of + over ⊕b
implies that the generating function of the pseudo-convolution is:
∞ j
!
M M
ĉ(z, b) = (di + ej−i ) + zj
b b
j=0 i=0
∞ j
!
M M
= (di + ej−i + zj)
b b
j=0 i=0
∞ j
!
M M
= (di + zi + ej−i + z(j − i)) . (192)
b b
j=0 i=0

By associativity, we can drop the large outer brackets and we can also change the summation order:

M ∞
M
ĉ(z, b) = (di + zi + ej−i + z(j − i)). (193)
b b
i=0 j=i

38

Electronic copy available at: https://ssrn.com/abstract=3929583


Using the distributivity of + over ⊕b :
 

M ∞
M
ĉ(z, b) = di + zi + (ej−i + z(j − i))
b b
i=0 j=i
∞ ∞
!
M M
= di + zi + (ek + zk) , (194)
b b
i=0 k=0

from re-indexing the inner sum. Again, using the distributivity of + over ⊕b , the generating function of
the pseudo-convolution is the sum of the generating functions:

M ∞
M
ĉ(z, b) = (di + zi) + (ek + zk)
b b
i=0 k=0
ˆ b) + ê(z, b).
= d(z, (195)

Just as a power series turns ordinary convolution into ordinary multiplication, our generating function
turns pseudo-convolution into ordinary addition.

References
[1] Aczel, J., 1966, Lectures on Functional Equations and their Applications, Academic Press, New York.
[2] Bachelier L., 1900, “Théorie de la spéculation.” Gauthier-Villars.
[3] Black F., and M. Scholes, 1973, “The Pricing of Options and Corporate Liabilities”, Journal of
Political Economy, May - Jun., 81, 3, pp. 637–654.
[4] Burgin M. and M. Czachor, 2020, Non-Diophantine Arithmetics in Mathematics, Physics and Psychology,
World Scientific.
[5] Carr P., and D. Costa, 2020, “Optionality as a Binary Operation”, NYU working paper.
[6] Carr P., and L. Torricelli, 2020, “Additive Logistic Processes in Option Pricing”, NYU working paper.
[7] Grossman, M. and R. Katz, 1972, Non-Newtonian Calculus.
[8] McEneaney W.M., 1999, “Exactly Linearizing Algebras for Risk-Sensitive Filtering”, Proceedings of
the Conference on Decision & Control, pp. 137–142.
[9] Pap E. 1993, “g-calculus”, Univ u Novom Sadu Zb Rad Prirod-Mat Fak Ser Mat 23:145-–156.
[10] Raı̈ssouli M. and I. Jebri, 2010, “Various Proofs for the Decrease Monotonicity of the Schatten’s Power
Norm, Various Families of Rn Norms and Some Open Problems”, Int. J. Open Problems Compt. Math
3, 2, pp. 164–174.
[11] von Helmholtz, H., 1882, Physical memoirs, selected and translated from foreign sources. Taylor &
Francis.

39

Electronic copy available at: https://ssrn.com/abstract=3929583

You might also like