1 Introduction to Arbitrage Pricing Theory For reference, this chapter reviews selected results from stochastic calculus and from the modern theory of asset pricing. The material in this chapter is well covered in existing literature, so we keep the chapter brief and the mathematical treatment informal. For a more rigorous treatment we refer to Duffie [2001] or Musicla and Rutkowski [1997]. Most of the necessary mathematical foundation for the theory is available in Karatzas and Shreve [1997], Oksendal [1992], and Protter (2005). The treatment in this chapter focuses on asset pricing in general; we shall specialize it to interest rate securities in Chapter 4. Chapter 5 introduces fixed income markets in detail. 1.1 The Setup Unless otherwise noted, in this book we shall always consider an economy with continuous and frictionless trading taking place inside a finite horizon {0,T]. We assume the existence of traded dividend-free assets with prices characterized by a p-dimensional vector-valued stochastic process X(t) = (X1(t),...Xp(t))'. Uncertainty and information arrival is modeled by a probability space (2,,P), with @ being a sample space with outcome elements w; F being a o-algebra on 92; and P being a probability measure on the measure space ({2, F). Information is revealed over time according to a filtration {F;, t € [0,T]}, a family of sub-o-algebras of F satisfying Fs © F, whenever s < t. We can loosely think of F; as the information available at time ¢. We assume that the process X(t) is adapted to {F;}, ie. that X(t) is fully observable at time t. For technical reasons, we require that the filtration satisfies the “usual conditions”+. Let EP(-) be the expectation \To satisfy the “usual conditions”, 7, must be right-continuous for all t, and Fo must contain all the null-sets of F, ie. all subsets of sets of zero P-probability.4 1 Introduction to Arbitrage Pricing Theory operator for the measure P; when conditioning on information at time t, we will use the notation EP (.) = EP (.|F2). Tn all of the models in this book, we specialize the abstract setup above to the situation where information is generated by a d-dimensional vector-valued Brownian motion (or Wiener process) W(t) = (Wi (t),-..,Wa(t))", where W; is independent of W; for i 4 j. Brownian motions are treated in detail in Karatzas and Shreve [1997]; here, we just recall that a scalar Brownian motion W; is a continuous stochastic process starting at 0 (ie. W;(0) = 0), having independent Gaussian increments: W;(t) -—W;(s) ~ N(0,t—s), € > s. The filtration we consider is normally always the one generated by W, F, = o{W(u), 0 < u < t}, possibly augmented to satisfy the usual conditions. We will generally assume that the price vector X(t) is described by a vector- valued Ito process: t t X(t) = X(0) +f L(s,w) as+ [ a(s,w)dW(s), (1.1) 0 0 or, in differential notation, dX (t) = w(t, w) dt + o(t,w) dW(t), (1.2) where jz: R x 2 -+ R? and a: R x 2 — R?*4 are processes of dimension p and p x d, respectively. We assume that both yz and o are adapted to {Fy} and are in L' and L? respectively, in the sense that for all t € (0,T], t f |n(s,w)| ds < 00, (1.3) t [ lo(s,w)|? ds < 00, (1.4) 0 almost surely”. In (1.4), we have defined Jo(t,w)|? = tr (o(t,w)o(t,w)"). (1.5) We notice that the sample paths of X generated by (1-1) are almost surely continuous, with no jumps in asset prices. A technical treatment of Ito processes and the Ito integral with respect to Brownian motion can be found in Karatzas and Shreve [1997]. For our needs, it suffices to think of the Ito integral as [ couyawi) = lim Soli )8,u) [W (id) —W((i-1)8)], (1.6) 5 ne isl ?An event holds “almost surely” — often abbreviated by “a.s.” — if the probability of the event is one.1.1 ‘The Setup 5 where 5 £ t/n. We note that the integrand ¢ is here always evaluated at the left of each interval {(i — 1)4, 16]. Other choices are possible?, but, as we shall see, the “non-anticipative” structure of the Ito integral gives rise to a number of useful results and makes it particularly useful as a model of trading gains (see Section 1.2). We list a few relevant definitions and results below. Definition 1.1.1 (Martingale). Let Y(t) be an adapted vector-valued pro- cess with EP (|Y (t)|) < 00 for all t € [0,T]. We say that Y(t) is a martingale under measure P if for all s,t € [0,7] witht < s, EP (¥(s))=Y(t), as. If we replace the equality sign in this equation with < or >, Y(t) is said to be a supermartingale or a submartingale, respectively. Definition 1.1.2 (Space H?). Let |o(t,w)|? be as defined in (1.5). We say that o is in H®, if for allt € [0,T] we have EP (f jo(s,w)|? as) < 00. The importance of Definition 1.1.2 becomes clear from the following result: Theorem 1.1.3 (Properties of Ito Integral). Define I(t) = Jo.o(s,w) dW(s) and assume that o is in H?. Then 1. I(t) is Fy-measurable. 2. I(t) is a continuous martingale. In particular, EP(I(t)) = 0 for all t¢ [0,7]. 3. EP(((t)|?) = EP(f, lo(s,w)|? ds) < 00. 4. EP(I(t)1(8)7) = EP (fen o(u,w)o(u,w)7 du). A proof of Theorem 1.1.3 can be found in, e.g., Karatzas and Shreve [1997]. The equality in the third item of Theorem 1.1.3 is known as the Ito isometry. Due to the inequality in the third item, we say that the martingale defined in the process is a square-integrable martingale. While it is common in applied work to simply assume that Ito integrals are martingales, without technical regularity conditions on o(t,w) (such as the H? restriction in Theorem 1.1.3), we should note that Ito integrals involving general processes in L? can, in fact, only be guaranteed to be local martingales. A process X is said to be a local martingale if there exists a 3Phe Stratonovich stochastic integral evaluates ¢ at the mid-point of each interval.6 1 Introduction to Arbitrage Pricing ‘Theory sequence of stopping times* {r, }°,, with tT, — co as n — oo, such that X(min(t,7)), > 0, is a martingale for all n. In other words, all “driftless” Ito processes of the type dY (t) = o(t,w) dW(t) (1.7) are local martingales, but not necessarily martingales. Interestingly, a con- verse result holds as well; all local martingales adapted to the filtration generated by the Brownian motion W can be represented as Ito processes of the form (1.7): Theorem 1.1.4 (Martingale Representation Theorem). Jf Y is a local martingale adapted to the filtration generated by a Brownian motion W, then there exists a process 0 such that (1.7) holds. If Y is a square-integrable martingale, then o is in H?. The proof of Theorem 1.1.4 can be found in Karatzas and Shreve [1997]. In the manipulation of functionals of Ito processes, the key result is a famous result by K. Ito: Theorem 1.1.5 (Ito’s Lemma). Let f(t,x), 2 = («1,...,2))', denote a continuous function, f :{0,T] x R? + R, with continuous partial derivatives Of /At = fr, Af [Oui = fo,, Pf /Ox,0x; = fr,x,. Let X(t) be given by the Ito process (1.2) and define a scalar process Y(t) = f(t, X(t)). Then Y(t) is an Ito process with stochastic differential dY (t) = f(t, X(t)) dt + fe (t, X(t) w(t,w) dt + fy (t, X (t)) o(t,w) dW (t) + ; SOD fave; (.X() (o(t.w)o(tew)"), 5 ty i=1 j=l where fz = (fr,,---5 fec,,)- For easy reference, the result below lists Ito’s lemma for the special case where p=d=1 Corollary 1.1.6. For the case p = d =1, Ito’s lemma becomes dy (t) (i (t, X(t)) + fe (X(t) ul(t,w) 4 plex (x(0) o(t4)?) dt + fa (t,X (8) o(t,w) dV(t). Ito’s lemma can be motivated heuristically from a Taylor expansion. For instance, for the scalar case in Corollary 1.1.6, we write informally Recall that a stopping time 7 is simply a random time adapted to the given filtration, in the sense that the event {r < t} belongs to Fy.1.2 ‘Trading Gains and Arbitrage 7 f (t+ dt, X(t + dt)) = f(t, X(t) + fedt + fedX(t) + phos (dX(0)" eee (1.8) Here, we have (dX (t))? = p(t,w)? (dt)? + o(t,w)? (dW(2))? + 2u(t,w)o(t,w) dt dW (t). As shown earlier, (dW (t))? = dt in quadratic mean, whereas all other terms in the expression for (dX (t))? are of order O(de?/?) or higher and can be neglected for small dt. In the limit, we therefore have (dX (t))? = o(t,w)? dt which can be inserted into (1.8). The result in Corollary 1.1.6 then emerges. Remark 1.1.7. The quantity (dX (t))? discussed above is the differential of the quadratic variation of X(t), often denoted by (X(t), X(t)). That is, d(X(t), X(t) = (dX)? = Le(. x10) = f (axeny?. For two different (scalar) Ito processes X(t) and Y(t), we may equivalently define the quadratic covariation process (X(t), ¥(t)) by (X(t), Y(t) = dX(t) dY (2). Sometimes we also write d(X(t),¥(t)) = (dX(t),d¥(t)). If X(t) is a p- dimensional process and Y(t) is a g-dimensional process, the quadratic covariation (X(t), Y(t)") is a (p x q)-dimensional matrix process whose (i, j)-th element is (X;(t), Y;(t)), 4=1,....p,7=1,...,¢. The so-called Tanaka extension (see Karatzas and Shreve {1997]) extends Ito’s lemma to continuous but non-differentiable functions. At points where the function has a kink, the Tanaka extension (loosely speaking) justifies using the Heaviside (step-) function for the first-order derivative and the Dirac delta function for the second-order derivative. An application of the Tanaka extension can be found in Section 1.9.2 and in Chapter 7, along with further discussion and references. 1.2 Trading Gains and Arbitrage Working in the setting of Section 1.1 with assets driven by Ito processes, we now consider an investor engaging in a trading strategy involving the p assets X),..., Xp. Let the trading strategy be characterized by a predictable® adapted process J(t,w) = (¢1(t,w),...,¢p(t,w))", with ¢;(t,w) denoting °A predictable process is one where we, loosely speaking, can “foretell” the value of the process at time t, given all information available up to, but not including, time t. All adapted continuous processes are thus predictable. For a technical definition of predictable processes, see Karatzas and Shreve [1997].8 1 Introduction to Arbitrage Pricing Theory the holdings at time ¢ in the i-th asset X;. The value (¢) of the trading strategy at time ¢ is thus (dropping the dependence on w in the notation) a(t) = o(t) X(t). (1.9) The gain from trading over a small time interval {t,t + 6) is (approximately) o(t)" [X(t + 4) — X(t)], suggesting (compare to (1.6)) that the Ito integral [ow T ax(s y= [os Tu(s)ads+ [ o(e)To (s)4W(s) is a proper model for trading gains over (0, ¢]. An investment strategy is said to be self-financing if, for any t € [0,7], ay —mo)= [ #(s)dX(s). (1.10) This relationship simply expresses that changes in portfolio value are solely caused by trading gains or losses, with no funds being added or withdrawn. Self-financing trading strategies allow investors to turn a certain initial investment 7(0) into stochastic future wealth z(t). Under natural assump- tions on possible trading strategies (e.g., that there is finite supply of all assets) we would expect that there should be limitations to the profits that self-financing strategies can create. Most notably, it should be impossible to create “something for nothing”, that is, to turn a zero initial investment into future wealth that is certain to be non-negative and may be positive with non-zero probability. To express this formally, we introduce the concept of an arbitrage opportunity: Definition 1.2.1 (Arbitrage). An arbitrage opportunity is @ self-financing strategy @ for which r(0) =0 and, for some t € [0,T), n(t)>0 as., and P (n(t) > 0) >0, (1.11) with m given in (1.9). In economic equilibrium, arbitrage strategies cannot exist and preclud- ing (1.11) constitutes a fundamental consistency requirement on the asset processes. 1.3 Equivalent Martingale Measures and Arbitrage We turn to the question of characterizing the conditions under which the trading economy is free of arbitrage opportunities. A concise way to state these conditions involves equivalent martingale measures, a concept we shall work our way up to in a number of steps. First, we recall that two1.3 Equivalent Martingale Measures and Arbitrage 9 probability measures P and P on the same measure space ({2,F) are said to be equivalent if P(A) = 0 + B(A) = 0, VA € F; that is, the two measures have the same null-sets. An important result from measure theory states that equivalent measures are uniquely associated through a quantity known as a Radon-Nikodym derivative: Theorem 1.3.1 (Radon-Nikodym Theorem). Let P and P be equivalent probability measures on the common measure space (2,F). There exists a unique (a.s.) non-negative random variable R with EP(R) = 1, such that P(A)=EP(R1yay), for all AEF. For a proof of Theorem 1.3.1, see e.g. Billingsley [1995]. The random variable R in the theorem is known as a Radon-Nikodym derivative and is denoted dP/dP. In the theorem we have used an indicator 1,4}; this quantity is 1 if the event A comes true, 0 if not. For later use, we associate any probability measure P with a density process s(t) =EP (s) » Wee (0,7). (1.12) Clearly, ¢(t) is a P-martingale with (0) = 1 and ¢(t) = EP (<(T)). A simple conditioning exercise demonstrates that for any Fp-measurable random variable Y(T), with R = dP/dP, EP(Y(D)|F) = ra (RY(T)F) = (t)'EP (EP(R|Fr)¥(T)|F.) oP) si) : (1.13) P Brn We shall use this result on numerous occasions in this book. We now introduce the important concept of a deflator, a strictly positive Ito process used to normalize the asset prices. Let the defla- tor be denoted D(t), and define the normalized asset process X?(t) = (Xi(t)/Dit),..., Xp(t)/D(t))™. We say that a measure Q? is an equivalent martingale measure induced by D if X(t) is a martingale with respect to Q?. If Q? is a martingale measure, we say that a self-financing trading strategy is permissible if t J 6s) ax(s) 0 is a martingale. For the Ito setup discussed earlier, a permissible strategy® is obtained by, say, requiring that (t)'o(t) is in H?; see Theorem 1.1.3. An °The technical restriction on trading positions imposed by only considering permissible trading strategies rules out certain pathological strategies, such as the10 1 Introduction to Arbitrage Pricing Theory application of Ito’s lemma combined with (1.9)—(1.10) implies that (t)/D(t) is a Q?-martingale when the trading strategy is permissible. For permissible trading strategies, the importance of equivalent martin- gale measures follows from the following theorem: Theorem 1.3.2 (Sufficient Condition for No-Arbitrage). Restrict at- tention to permissible trading strategies. If there is a deflator D such that the deflated asset price process allows for an equivalent martingale measure, then there is no arbitrage. For a proof we refer to Musiela and Rutkowski [1997]. We note that Theorem 1.3.2 only provides sufficient conditions for the absence of arbitrage, and known (and rather technical) counterexamples demonstrate that the existence of an equivalent martingale measure does not follow from the absence of arbitrage in a setting with permissible trading strategies. A body of results known as the fundamental theorem of arbitrage establishes the conditions under which the existence of an equivalent martingale measure is also a necessary condition for the absence of arbitrage. The results are rather technical, but generally state that absence of arbitrage and the existence of an equivalent martingale measure are “nearly” equivalent concepts. The exact notion of “nearly” equivalent is discussed in Duffie [2001] as well as in the authoritative reference’ Delbaen and Schachermayer [1994]. For our purposes in this book, we ignore many of these technicalities and often simply treat the absence of arbitrage and the existence of a martingale measure as equivalent concepts. Finally, if the deflator is one of the p assets, we call the deflator a numeraire. Let us, say, assume that X, is strictly positive and can be used as a numeraire. Also assume that a deflator D has been identified such that Theorem 1.3.2 holds. As X;(t)/D(t) is a Q?-martingale, we can use the Radon-Nikodym theorem to define a new measure Q*? by the density s(t) = (Xi (t)/D(t))/(X1(0)/D(0)). For an Fyp-measurable variable Y(T), we then have, from (1.13), xyoe?” (262) = vioee” (53). (4.14) In particular, if Y(t)/D(t) is a Q?-martingale, Y(t)/X1(t) must also be a Q*:-martingale. In practice, it normally suffices to only consider deflators from the set of available numeraires. Remark 1.3.3. Some sources define 1/D(t) (rather than D(t)) as the deflator. The convention used in this book is more natural for our applications. doubling strategy considered in Harrison and Kreps [1979]. A realistic resource- constrained economy will always bound the size of the positions one can take in an asset, sufficing to ensure that predictable trading strategies are permissible. "In a nutshell, Delbaen and Schachermayer [1994] show that absence of arbitrage implies only the existence of a local martingale measure.1.4 Derivative Security Pricing and Complete Markets il 1.4 Derivative Security Pricing and Complete Markets A T-maturity derivative security (also known as a contingent claim) pays out at time T an Fr-measurable random variable V(T), and makes no payments before T. We assume that V(T) has finite variance, and say that the derivative security is attainable (or sometimes redundant) if there exists a permissible trading strategy ¢ such that V(T) = ¢(T)' X(T) = x(T) as. The trading strategy is said to replicate the derivative security. Importantly, the absence of arbitrage dictates that the time 0 price of an attainable derivative security V(0) must be equal to the cost of setting up the self- financing strategy, i.e. V(0) = (0). More generally, V(t) = x(t), t € (0, T]. This observation is the foundation of arbitrage pricing and allows us to price derivative securities as expectations under an equivalent martingale measure. Specifically, consider a deflator D and assume the existence of an equivalent martingale measure Q? induced by D; the existence of Q? guarantees that there are no arbitrages in the market, by Theorem 1.3.2. Now, from the martingale property of 7(t)/D(t) in the measure Q? and the relation V(t) = 7(t) it immediately follows that we? (UM) Dit) ~* \D(T) or V(t) = D(t)EO” (FR) ; (1.15) If all finite-variance #7-measurable random variables can be replicated, the market is said to be complete. In a complete market, all derivatives are “spanned” and hence have unique prices. Interestingly, a similar uniqueness result holds for equivalent martingale measures: Theorem 1.4.1. In the absence of arbitrage, a market is complete if and only if there exists a deflator inducing a unique martingale measure. From (1.14) it follows that the martingale measures induced by all numeraires must then be unique as well. In practical applications, we shall often manipulate the choice of nu- meraire asset to simplify computations. The following result is useful for this: Theorem 1.4.2 (Change of Numeraire). Consider two numeraires N(t) and M(t), inducing equivalent martingale measures QN and Q™, respectively. If the market is complete, then the density of the Radon-Nikodym derivative relating the two measures is uniquely given by iy paw (dQ!) _ M(t)/M(0) o=88 (Sr) = Mam12 1 Introduction to Arbitrage Pricing ‘Theory Proof, As the market is complete, all derivatives prices are unique. Consider an integrable Fr-measurable payout V(T) = Y(T)M(T), with time ¢ price V(t). From Theorem 1.4.1 and (1.15) we must have V(t) = N()ES" (on?) = M(t)ER" (“On-) or Be" (y(r) = £2" ( vt MEINE, qT) ) M(t)/N(t) Comparison with (1.13), and the fact that the density must be scaled to equal 1 at time 0, reveals that the Radon-Nikodym derivative for the measure shift is characterized by the density in the theorem. 0 1.5 Girsanov’s Theorem The last two sections have demonstrated a close link between the concept of arbitrage and the existence and uniqueness of equivalent martingale measures. In this section, we consider i) the conditions on the asset prices that allow for an equivalent martingale measure; and ii) the effect on asset dynamics from a change of probability measure. We consider two measures P and P(@) related by a density <°(t) = EP (dP(@)/dP), where <8(t) is an exponential martingale given by the Ito process dc? (t)/s0(t) = —0(t) dW (t), where W(t) is a d-dimensional P-Brownian motion. The d-dimensional process @ is known as the market price of risk. By an application of Ito’s lemma, we can write <(t) = exp (- [ar dW(s) - if 8s)" 8(s) a) 26g -[ 6(s)7 aiv(s)) (1.16) where &(-) is the Doleans exponential. An often-quoted sufficient condition on 6(t) for (1.16) to define a proper martingale (and not just a local martingale) is the Novikov condition EP [exw (5 [asrases)| < 00. (1.17) The Novikov condition can often be difficult to verify in practical applications. Armed with the notation above, we are now ready to state the main result of this section.1.5 Girsanov’s Theorem 13 Theorem 1.5.1 (Girsanov’s Theorem). Suppose that <®(t) defined in (1.16) és @ martingale. Then for all t € (0,T] W(t) = w+ fm) As is a Brownian motion under the measure P(6). To discuss a strategy to prove Girsanov’s theorem, assume for simplicity that the dimension of the Brownian motion is d = 1. One way to construct a proof for Theorem 1.5.1 is to demonstrate that the joint moment. -generating function (mgf)® (under P(@)) of the increments W(t), W? (ta) ~ W(t), W8(tn) —W(tnay O< tr (Se. (W(t) — W(t »))-] Tle? (ti — ti-1) /2), 1 where we have defined to = 0. While carrying out such a proof is not difficult, we here merely justify the final result by examining the case n = 1 only. Specifically, we consider EP [exp (aW9(t))], where a € R and t > 0. Shifting probability measure, we get 1 EPO) [exp (aWv%(t))] = EP [exp (owe) 4a [ 0s) as)| oO =EP [exe (awe + af as) as) é (-[ 4(s) awis))| = et /2pP [ex (f (a — 6(s)) dW(s) — ; [i — %3))2as) | = ot/2pP le ([ (a —6(s)) awis))| 2, =e «2 Recall that the moment-generating function of a random variable Y in some measure P is defined as the expectation E?(exp(aY)), a € R. Unlike the charac- teristic function, the moment-generating function is not always well-defined for all values of the argument a.14 1 Introduction to Arbitrage Pricing ‘Theory as desired. In the last step, we used the fact that the Doleans exponential is a martingale with initial value 1. Girsanov’s theorem implies that we can shift probability measure to transform an Ito process with a given drift to an Ito process with nearly arbitrary drift. Specifically, we notice that our asset price process (under P) dX (t) = p(t) dt + o(t) dW(t) can be written AX (t) = (u(t) — o(t)6(t)) dt + o(t)dW(2), where W®(t) is a Brownian measure under the measure P(9). This process will be driftless provided that 6 satisfies the “spanning condition” s(t) = o(t)0(t) for all t € [0,7]. This gives us a convenient way to check for the existence of equivalent martingale measures: Corollary 1.5.2. For a given numeraire D, assume that the deflated asset process satisfies dX? (t) = p(t) dt + 0? (t) dW(t), where o?(t) is sufficiently regular to make fy o?(s)dW(s) a martingale. Assume also that there exists a @ such that the density ¢® is a martingale d (as.) 7 oP (t)a(t) = w(t), t€ (0,7), (1.18) then D induces an equivalent martingale measure and there is no arbitrage. Equation (1.18) is a system of linear equations and we can use rank results from linear algebra to determine the circumstances under which (1.18) will have solutions (no arbitrage) and when these are unique (complete market). For instance, a necessary condition for the market to be complete is that rank(o) = d. Further results along these lines can be found in Musiela and Rutkowski [1997] and Duffie [2001]. We conclude this section by noting that while a change of probability measure affects the drift 4 of an Ito process, it does not change the diffusion coefficient ¢. This is sometimes known as the diffusion invariance principle. 1.6 Stochastic Differential Equations So far we have defined the asset process vector to be an Ito process with general measurable coefficients j1(t,w) and o(t,w). In virtually all applica- tions, however, we restrict our attention to the case where these coefficients1.6 Stochastic Differential Equations 15 are deterministic functions of time and the state of the asset process®. In other words, we consider a stochastic differential equation (SDE) of the form dX(t) = p(t, X(t) dt +0 (t, X(t) dwt), X(0) = Xo, (1.19) with ys: [0,7] xR? > R?; 0 : (0,7] xR? - R°*4; and Xo an initial condition. A strong solution'® to (1.19) is an Ito process t x= Xo [ wlax(s) as | o(s,X(s)) dW(s). 0 A number of restrictions on jz and ¢ are needed to ensure that the solution to (1.19) exists and is unique. A standard result is listed below. Theorem 1.6.1. In (1.19) assume that there exists a constant K such that for allt € [0,T| and all x,y € R?, \u(t,2) ~ plt,y)| + fol) —o(t,y)| < Kix — yl, (Lipschitz condition), u(t, x)|? + |o(t,x)|? < K? (1+ |2/?), (growth condition). Then there exists a unique solution to (1.19). We notice that the dynamics of (1.19) do not depend on the past evolution of X(t) beyond the state of X at time t. This lack of path-dependence suggests that X is a Markov process. We formalize this as follows. Definition 1.6.2 (Markov Process). The R?-valued stochastic process X(t) is called a Markov process if for all s,t € [0,T] witht Ve, (t) us (t) dt i=1 + ; > Ve,x,(t)3i,5(t) dt + » Ve, (t)o;(t) dW(t), (1.24) where @; is the i-th row of the px d matrix o and Yj; is the (i, j)-th element in oo”. We recall that subscripts like Vy, denote partial differentiation, see Theorem 1.1.5. If V(t) can be replicated by a self-financing trading strategy ¢ in the p assets, we must also have, from (1.10), P P V(t) = ot) AX(t) = S> dilt)ua(t) dt + > gilt)oi(t) a(t). (2.28) Comparing terms in (1.24) and (1.25) we see that both equations will hold, provided that for all t € [0,7] av (t, X(t)) di(t) = an (1.26) and P 1 j=l To the extent that the system above allows for a solution (it may not if the market is not complete, from (1.26) we see that the trading strategy that replicates the derivative V holds V(t, X(t))/Ax; units of asset X; at time18 1 Introduction to Arbitrage Pricing ‘Theory t. The quantity 0V/dz; is often known as the delta with respect to X;!". Note that, from (1.9) and (1.26) we have that V@X() = MEAD x, (1.28) i=l Besides identifying an explicit replication strategy, the arguments above have also produced (1.27), a partial differential equation (PDE) for the value function V(t,2). The PDE is a second-order parabolic equation in p spatial variables, with known terminal condition V(T,x) = g(x) (a so-called Cauchy problem). Solving this PDE provides an alternative way to price the derivative, as compared to the purely probabilistic expectations-based methods outlined earlier (see (1.15)). We shall investigate the link between expectations and PDEs in more detail in Section 1.8. Inspection of the valuation PDE (1.27) reveals that the drifts jz; of the asset price SDE (1.19) are notably absent, making the price of the derivative security independent of drifts. This is typical of derivatives in complete markets and follows from the fact that derivatives can be priced preference- free, by arbitrage arguments. In contrast, for the elements of the fundamental asset price vector, risk-averse investors would demand that assets with high volatilities |o;| be rewarded with higher drifts (more precisely, higher rates of return) as compensation for the additional uncertainty. 1.8 Kolmogorov’s Equations and the Feynman-Kac Theorem In earlier sections, we have seen that derivatives prices can be expressed as expectations under certain probability measures or as solutions to PDEs. This hints at a deeper connection between expectations and PDEs, a connection we shall explore in this section. As part of this exploration, we list results for transition densities that will be useful later in model calibration. As in Section 1.6, we consider a Markov vector SDE of the type (see (1.19)) dX(t) = n(t, X(t) dt +0 (t, X(t) W(t), X(0) = Xo, (1.29) where the coefficients are assumed smooth enough to allow for a unique solution (see Theorem 1.6.1). Now define a functional u(t,x) = EP (g(X(T))|X() = 2), “Note that taking a position in V and following a trading strategy with $: = —OV/Ax;, i = 1,...,p will effectively remove any exposure to V (as we simultaneously take a long position in V and, through a trading strategy, a short position in V). ‘This strategy is known as a delta hedge.1.8 Kolmogorov’s Equations and the Feynman-Kac Theorem 19 for a function g : R? + R. Under regularity conditions on g, it is easy to see that the process u(t, X(t)), being a conditional expectation, must be a martingale. Proceeding informally, an application of Ito’s lemma gives, for t € [0,T) (suppressing dependence on X(t)), P PP du(t) = u(t) dt + 37 ue, (t)unr(t) ash SOS tere, (t)2i,9 (dt + O(AW(t)), i=l t=1 j=1 where as before ¥j,; is the (7, 7)-th element of oo’. From earlier results, we know that for u(t, X(¢)) to be a martingale, the term multiplying dt in the equation above must be zero. Defining the operator P a P ar A= dou eae +5 LL Sula ae 1 j=l we deduce that u(t, x) satisfies the PDE Ou(t, x) at Nie Ms i + Au(t,x) = 0, (1.30) with terminal condition u(T, a) = g(x). The equation above is known as the Kolmogorov backward equation for the SDE (1.29). The operator A is known as the generator or infinitesimal operator of the SDE, and can be identified as EP (u(t +h, X(t+ h)) |X(t) =e) — ult, x) ——— In arriving at (1.30) we made several implicit assumptions, most notably that the function u(t, x) exists and is twice differentiable. Sufficient conditions for the validity of (1.30) can be found in Karatzas and Shreve [1997], for instance. A relevant result is listed below. Au(t,2) = kim Theorem 1.8.1. Let the process X(t) be given by the SDE (1.29), where the coefficients 4. and o are continuous in x and satisfy the Lipschitz and growth conditions of Theorem 1.6.1. Consider a continuous function g(x) that is either non-negative or satisfies a polynomial growth condition, meaning that for some positive constants K and q g(x) < K(1+|a|%), «Ee R?. If u(t,x) solves (1.30) with boundary condition u(T,x) = g(x), and u(t, x) satisfies a polynomial growth condition in x, then u(t,2) = EP (g(X(T))|X()=2), t€ (0,7). (1.31) Conditions required to ensure existence of a solution to (1.30) are more involved, and we just refer to Karatzas and Shreve [1997] and the references therein.20 1 Introduction to Arbitrage Pricing ‘Theory A family of functions g of particular importance to many of our applica- tions is 7 g(a) =e *, kER?, where i = V—I is the imaginary unit. In this case u(t,z) becomes the characteristic function of X(T), conditional on X(t) = x. We refer to any standard statistics textbook (e.g. Ochi [1990]) for the many useful properties of characteristic functions. For the Markov process X(t) in (1.29), let us now introduce a transition density, given heuristically by P(t, 3 8,y) dy = P(X(s) € [y,y+dyl|X(t)=2), OSt t from a known state at time t, rather than vice-versa. For this, we first define an operator A* by Pp aflei)=-3 Sols you) re i= & (Ei 5(s, y)f(s,9)} Xu Ouidu, , In the transition density p(t,x;s,y) now consider (t,x) fixed and let A* operate on the resulting function of s and y. Under additional regularity conditions, we then have the forward Kolmogorov equation a; = Sels.9) + A*p(s,y) =0, (t,x) fixed, (1.33) subject to the boundary condition p(t, x;t,y) = 6(2 — y). The forward Kolmogorov equation is sometimes known as the Fokker- Planck equation. We stress that the backward equation is more general than the forward equation, in the sense that the former holds for general terminal conditions g(x), whereas the latter only holds for 5-type initial conditions.1.9 Black-Scholes and Extensions 21 We round off this section by a useful extension to the Kohnogorov backward equation. Specifically, consider extending the PDE (1.30) to Ou(t, x) ot + Au(t,a) + h(t,x) =r(t,a)u(t,2), (1.34) where h,r : [0,7] x R? > R. Given the boundary condition u(T,2) = g(x), the Feynman-Kac solution to (1.34), should it exist, is given by X(t)= °) ; (1.35) T u(t, 2) = EP (worn (X(P)) +/ w(t, s)h(s, X(s)) ds where . w(t,T) = exp (-/ ra.X(0) a), te (0,7), t The result is easily understood from an application of Ito’s lemma, similar to the one used above to motivate the backward Kolmogorov equation. Sufficient regularity conditions for the Feynman-Kac result to hold are identical to those of Theorem 1.8.1, supplemented with the requirement that r be nonnegative and continuous in «; and the requirement that h be continuous in x and either be nonnegative or satisfy a polynomial growth requirement in x. See Duffie [2001] for further details about the often delicate regularity issues surrounding the Feynman-Kac result. For later use, let us finally note that when g(x) = 6(a—y) and h(t, x) = 0, u(t, x) in (1.35) will equal Glt,a;Tyy) £ EP (e° WX AG (X(T) —y)|X() = 2). The function G is known as a state-price density or as an Arrow-Debreu security price function. In particular, notice that for an arbitrary g(x), we then have BP (eI MOG x (TY) XC) =2) = f Clea: T valu) dy. (1.86) Comparison with (1.32) shows that the state-price density is, essentially, equivalent to a Green’s function with built-in discounting. 1.9 Black-Scholes and Extensions In reviews of asset pricing theory, a discussion of the seminal Black-Scholes- Merton model (sometimes just known as the Black-Scholes model) of Black and Scholes [1973] and Merton [1973] is nearly mandatory. As the Black- Scholes-Merton (BSM) model constitutes a well-behaved setting in which to tie elements of previous sections together, our text is no exception. To provide a smoother transition to material that follows, we do, however, extend the usual analysis to include a simple case of stochastic interest rates.22 1 Introduction to Arbitrage Pricing Theory 1.9.1 Basics In the basic BSM economy, two assets are traded: a money market account B and a stock S. In previous notations, X(t) = (B(t), S(t)" and p = 2. The money market account value is 1 at time 0 and accrues risk-free interest at a continuously compounded, non-negative rate of r, initially assumed constant. The dynamics for § are thus given by an ordinary differential equation (ODE) dB(t)/A(t)=rdt, B(0)=1, implying that simply A(t) = B(0)e"'. The stock dynamics are assumed to satisfy GBMD under measure P: dS(t)/S(t) = pdt +odW(t), (1.37) where W is a Brownian motion of dimension d = 1, and y and o are constants. Taking first a probabilistic approach, we notice that f is positive and can be used as a numeraire. Let $°(t) = S(t)/A(t) be the stock price deflated by 6. By Ito’s lemma, d$9(t)/$9(t) =(u—r)dt +oaW(t). Applying Girsanov’s theorem (see Theorem 1.5.1) and Corollary 1.5.2, we see that if o 4 0, 6 will induce a unique equivalent martingale measure, with the measure shift characterized by the density process!? ds(t)/s(t)=-0dW(t), O= moe Clearly, ¢(t) defines an exponential martingale. The probability measure induced by the money market account £ is called the risk-neutral martingale measure and is traditionally denoted Q. Under Q, W4(t) = W(t) + 6t is a Brownian motion, and dS®(t)/S*(t) =o dW*(t), dS(t)/S(t) =rdt + odW*(t), (1.38) or, from (1.21), S(T) = S(t)e FTO 4o(WPT)-WPO), +t € [0,7]. (1.39) We note that under Q, the drift of the stock process is replaced by the risk-free interest rate r. That is, under Q agents in the economy will 12'The reader may recognize the market price of risk 9 as the Sharpe ratio of the stock S, a measure of how well the risk of stock (represented by a) is compensated by excess return (represented by —r).1.9 Black-Scholes and Extensions 23 appear to be indifferent (“neutral”) to the risk of the stock, content with an average growth rate of the stock equal to that of the money market account. Before proceeding with the BSM analysis, we wish to emphasize that the drift restriction imposed on the stock in the risk-neutral measure Q is a general result. In a larger setting with a p-dimensional vector asset process X, if the Q-dynamics of the components of X are all of the form dX;(t) =rXidt+O(dW(t)), i=1,...,p, there is no arbitrage. This result holds unchanged if the interest rate is random (see Section 1.9.3). Returning to the BSM setting, we note that the risk-neutral measure is unique, whereby the market is complete and all derivative securities on S (and #) are attainable. Let us consider a few such securities. First, we consider a security paying at time T $1 for certain. Such a security is a discount bond and we shall denote its time t price by P(t,T), t € (0,7). If the interest rate is positive, we would expect P(t,T) <1 as a reflection of the time value of money, with equality only holding for t = T. Application of the basic derivative pricing equation (1.15) immediately gives 1 P(t,T) = B(t)Ee (a) =E? (err) =en(T-), This result is trivial, as it is easily seen that the amount e-T(T-9 invested in the money market account at time t will grow to exactly $1 at time T. Second, consider a derivative V paying V(T) = S(Z’) — K at time T, with K being an arbitrary constant. Proceeding as above, at time t K, the call option is in-the-money (ITM) and the put option is out-of-the-money (OTM). If S(t) < K, the call is OTM and the put is ITM. The ATM, ITM, and OTM monikers are sometimes used to refer to the ordering of the forward value E,(S(T)) = S(t)e7-9 (for a T-maturity option) rather than the spot S(t), relative to the strike K. In deriving (1.43), the choice of $ as numeraire was arbitrary. If we instead use S (which is also strictly positive) as numeraire, we can write (SZ) — K)* ao) = S108" ( 30) ) = see" (a= K/s()*), (1-44) where Q is the martingale measure induced by S. To identify the measure shift involved in moving from P to QS, consider that 85(t) = 8(t)/S(t) must be a martingale in Q®. By Ito’s lemma, in measure P we have dB5(t)/BS(t) = (r — w+ 0?) dt — o dW(t), such that dW5(t) = ((r —)/o + @) dt + dW/(t) is a Brownian motion under Q°. Application of Ito’s lemma on 1/S(t) yields, after a few rearrangements, dS(t)“*/S(t)7! = -rdt — 0 dWS(t),1.9 Black-Scholes and Extensions 25, which is a GBMD as before. Evaluation of the expectation (1.44) can be verified to recover the BSM formula (1.43). Our derivation of the BSM formula was so far entirely probabilistic. Writing c(t) = c(t, 8, $), the arguments in Section 1.7 allow us to write c as a solution to the PDE (see (1.27)) (1.45) subject to the boundary condition ¢(T, 6, 8) = (S—K)*. From (1.28) we also have that the replication positions in B and S are 2& and 2&, respectively. That is, Oc Oc t,B,S) ==, =o8. 1.46) (8,8) = FEB + 55 (1.46) As 8 is deterministic, we can actually eliminate c-dependence on this variable by a change of variables ¢(t, S) = c(t, 8,S). By the chain rule de Oc IcAB Ac , Ae, Ac dc, a topo at op a tT ag” where the last equation follows from (1.46). Inserting this into (1.45) yields the original Black-Scholes PDE Oe 8 lop PF ~ at Sag + 9% Ss age 7% (1.47) with @T, S) = (S — K)*. We can solve this equation by classical methods (see Lipton [2001] for several techniques), or we can use the Feynman-Kac result to write it as an expectation. We leave it as an exercise to the reader to verify that Feynman-Kac leads to the same expectation as derived earlier by probabilistic means (see (1.41)). A final note: the derivation of the Black-Scholes PDE above was somewhat, non-standard due to the initial assumption of option price being a function of the deterministic numeraire 8. A more conventional (but entirely equivalent) argument sets up a portfolio of the call option and a position in the stock, and demonstrates that the stock position can be set such that the total portfolio growth is deterministic (risk-free) on [t,t+dt]. Equating the portfolio growth ate yields the Black-Scholes PDE (1.47). See Hull {2006} is approach. 1.9.2 Alternative Derivation We have already demonstrated several different ways of proving the BSM call pricing formula, but as shown in Andreasen et al. {1998] there are many more. One particularly enlightening proof is based on the concept of local time and shall briefly be discussed in this section. The proof, which borrows26 1 Introduction to Arbitrage Pricing ‘Theory from the results in Carr and Jarrow [1990], will also allow us to demonstrate the Tanaka extension of Ito’s lemma, mentioned earlier in Section 1.1. As above, we assume that the stock price process is as in (1.38), and define the forward stock price F(t) = S(t)/P(t,T). Clearly, dF(t)/F(t) =o dW(t), t«} and the second derivative can be interpreted as the Dirac delta function, 5(F(t) — K). As I is clearly not twice differentiable, Ito’s lemma formally does not apply, but the Tanaka extension nevertheless gives us permission to write al(t) = Aprqsmy APO + 50? FUe)?6 (F(A) — KY) = ltra>Ky oF (t) Wht) + 5o7K25 (F(t) —K) dt. In integrated form, ‘T 1 T I(T) = I(t) +f Lpe(uy>K} OF (u) dW? (u) + so? f 5(F(u) — K) du. t t The second integral in this expression is a random variable known as the local time of F spent at the level KX, on the interval {t, T}. Taking expectations, it follows that Ee l op2 [ pa ¢ (U(T)) = T(t) + 50° [ E? (5 (F(u) — K)) du. 2 t Here, if p(t,y;u,x) is the density of F(u) given F(t) = y, u > t, then obviously Be (5 (F(u) — K)) =p(t, F()su, K). By the definition of F(T) we have F(T) = S(T), such that I(T) = (S(T) — K)*. From (1.41), we may therefore write the time ¢ European call option price as e(t) = P(t, T)EP (I(T)) ‘T = (S(t) -— KP(t,T))* + PED 21? | p(t, F(t);u,K)du. (1.49) A The formula (1.49) decomposes the call option into a sum of two terms, the intrinsic value and the time value, respectively. The time value can be made more explicit by observing from the representation (1.39) that!® This also follows directly from the fact that F(u) is a log-normal random variable with moments given by (1.22) and (1.23).1.9 Black-Scholes and Extensions 27 ou Ky = ex (hates P(t PO: K) = pee »( alu) , d(u) & MEO [30 — ot =) In other words, we have arrived at the following result. Proposition 1.9.3. The European call option price c(t) on the process (1.38) can be written as 7 7 . P(t, T)oK [7 $(d(u)) . a(t) = (810) — KET)? + FODEE f SO au, (1.50) where (x) is the Gaussian density. Explicit evaluation of the integral in (1.50) can be verified to produce the BSM formula in Theorem 1.9.1. We leave this as an exercise to the reader. 1.9.3 Extensions [Link] Deterministic Parameters and Dividends In our basic BSM setup, consider now first a simple extension to a deter- ministic interest rate r(t) and a deterministic volatility a(t). Carrying out the analysis as before, we see that discount bond prices now become P(t,T) = ef r)as, (1.51) The BSM call pricing formula (1.43) holds unchanged provided P(t, T) is changed according to (1.51), and we redefine a, & R(S@)/K) + A (r(s) = 0(s)?/2) ds LTS —aoooooe Kr a(s)? ds Let us further assume that the stock pays dividends at a deterministic rate of q(¢). Our framework so far, however, has assumed that assets pay no cash over [0,7]. To salvage the situation, consider a fictitious asset S* obtained by reinvesting all dividends into the stock S itself. It is easily seen that S*(t) = S(t)eli a9) as, and clearly S*(t) satisfies the requirements of generating no cash flows on [0,7]. Stating the call option payout as e(T) = (S(T) — K)t = (s*(ryem Jor afs) ds _ x)”28 1 Introduction to Arbitrage Pricing Theory and performing the pricing analysis of Section 1.9.1 on S*(t), rather than S(t), results in a dividend-extended BSM call option formula: c(t) = S(t)en 9) 486(d,) — K P(t, T)#(d_), dg & MUSIK) + fr (r(s) — 4(s) + (s)?/2) ds : Vf" o(s)? ds When the stock pays a dividend rate of q(t), note that the risk-neutral process for S(t) is AS(t)/S(t) = (r(t) ~ a(t) dt + o(t) awe), which extends (1.38). Note that for the special case where r(t) = q(t), S(¢) becomes a martingale and the call option price formula simplifies to c(t) = P(t, T) (S(t)®(d,) — K@(d_)), (1.52) where now a In(S(t)/K) £3 f7 o(s)2ds 4, 2 Se es gr o(s)? ds d. Remark 1.9.4. The martingale call formula (1.52) typically emerges when pricing options on futures and forward prices (see (1.48)) and is often called the Black formula, in honor of the work in Black [1976]. [Link] Stochastic Interest Rates We now get, even more ambitious and wish to consider call option pricing in the case where the interest rate r is stochastic. The money market account B becomes B(t) = ef (9) 48, and is now assumed an F;-measurable random variable. Proceeding as in Section 1.9.1, we find that under the risk-neutral measure Q, the call option price expression is (assuming that the stock pays no dividends) 1 T (t) = B()ES ( —— (S(T) — K)* ) = ES (eW fr) as ~K)*). a(t) = BOWES ( ser (S(T) KY") = 88 (eM 04 (9(2) ~ KY") (1.53) In (1.53), we emphasize that the numeraire no longer can be pulled out from the expectation. Still, to simplify call option computations, it would be convenient to somehow remove the term exp(— fi r(s) ds) from the expectation in (1.53). By substituting 1 for ($(7) — K)+ in the expression above, we first notice that1.9 Black-Scholes and Extensions 29 PT) = BP (ek rede) | This inspires us to perform a new measure shift, where we use the discount bond P(t, T), rather than 6(t), as our numeraire. Let the martingale measure induced by P(t,T’) be denoted Q?, often termed the T-forward measure. By the standard result (1.15) we have elt) = PU TIES” (P(T,T)" (S(L) - K)*) = PU TIES” ((S(2) - K)*), where we have used that P(T,T) = 1. From Theorem 1.4.2, Q? and Q are related by the density \T P. To proceed, we need to add more structure to the model by making assumptions about the stochastic process for P(t,T). We shall spend con- siderable effort in subsequent chapters on this issue, but for this initial application we simply assume that P(t,T’) has Q dynamics dP(t,T)/P(t,T) = r(t) dt — op(t,T) dWr(t), (1.55) where op(t,T) is deterministic and Wp(t) is a Brownian motion correlated to the stock Brownian motion. Notice that the drift of P(t,T) under Q is not freely specifiable and must be equal to the risk-free rate; see the discussion following (1.38). For clarity, let the stock Brownian motion be renamed Ws(t), and assume that the correlation between Wp(t) and Ws(t) is a constant p. In the setting of vector-valued Brownian motion with independent components used in earlier sections, we can introduce correlation by writing W(t) = (Wi(t), Wo(t))" and setting, say, W(t) = Wilt), Ws(t) = pWi(t) + V1 — p>Walt). The filtration {F;,} of our extended BSM setting is the one generated by the 2-dimensional W(t). Under Q’, the deflated process S?(t) = S(t)/P(t,T) is a martingale. An application of Ito’s lemma combined with the Diffusion Invariance Principle shows that the Q? process for S?(t) is dS? (t)/S? (t) = op(t,T) dWi(t) + a(t) (paws (t) + + i — pe dWo(t) ), (1.56) where a(t) as before is the deterministic volatility of the stock S. We recognize S(t) as a drift-free geometric Brownian motion with instantaneous variance of30 1 Introduction to Arbitrage Pricing ‘Vheory op(t,T)? + 0(t)? + 2po(t)op(t.T). Exploiting the convenient fact that S?(T) = S(T) and c(T) = (S?(T)—K)* (as P(T,T) = 1), we get oft) = P(t, TBS” ((S°(2) - K)*) sa} + = pur) [ (SP (yer drenes VO?) - K) o(z)dz, (1.57) 00 where we have defined the “term”, or total, variance T 9 v(t,T) & [ (op(s,T)? + o(s)? + 2p0(s)or(s,T)) ds. (1.58) Je Completing the integration (compare with (1.42)) and using S?(t) = S(t)/P(t,T), we arrive at a modified BSM-type call option formula: Proposition 1.9.5. Consider a BSM economy with stochastic interest rates evolving according to (1.55). Define term variance v(t, T) as in (1.58). Then, the T-maturity European call option price is lt) = S)O(dy) — KPULT)OW.), _ In(S(t)/ (KPT) = 30(t,T) u(t,T) : Proposition 1.9.5 was originally derived in Merton [1973], using PDE methods. Extensions to dividend-paying stocks are straightforward and follow the arguments shown in Section [Link]. ds 1.10 Options with Early Exercise Rights In our previous definition of a contingent claim, we assumed that the claim involved a single F-measurable payout at time T. In reality, a number of derivative contracts may have intermediate cash payments from, say, scheduled coupons or through “rebates” for barrier-style options. Mostly, such complications are straightforwardly incorporated; see for instance Section 2.7.3. Of particular interest from a theoretical perspective are the claims that allow the holder to accelerate payments through early exercise. Derivative securities with early exercise are characterized by an adapted payout process U(t), payable to the option holder at a stopping time (or exercise policy) tT < T, chosen by the holder. If early exercise can take place at any time in some interval, we say that the derivative security is an American option; if exercise can only take place on a discrete set of dates, we say that it is a Bermudan option.1.10 Options with Early Exercise Rights 31 Let the allowed (and deterministic) set of exercise dates larger than or equal to t be denoted D(t), and suppose that we are given at time 0 a particular exercise policy 7 taking values in D(0), as well as a pricing numeraire NV inducing a unique martingale measure Q%. Let V7(0) be the time 0 value of a derivative security that pays U(r). Under some technical conditions on U(t), we can write for the value of the derivative security V7(0) = BE” (Aa) , (1.59) where we have assumed, with no loss of generality, that N(0) = 1. Let T(t) be the time ¢ set of (future) stopping times taking value in D(t). In the absence of arbitrage, the time 0 value of a security with early exercise into U must then be given by the optimal stopping problem w (U(r) V(0)= sup V7(0)= sup E® ( ) : (1.60) reT(0) reT(0) N(x) reflecting the fact that a rational investor would choose an exercise policy to optimize the value of his claim. We can extend (1.60) to future times ¢ by Hon gaye (HB). where supe7y) E2’ (U(r)/N(r)) is known as the Snell envelope of U/N under Q". The process V(t) must here be interpreted as the value of the option with early exercise, con mal on exercise not having taken place before time t. To make this explicit, let 7* € T(0) be the optimal exercise policy, as seen from time 0. We can then write, for0 E%” (V(t)/N()), which establishes that V(t)/N(t) is a supermartingale under Q’. This result also follows directly from known properties of the Snell envelope; see, e.g., Musiela and Rutkowski [1997]. For later use, focus now on the Bermudan case and assume that D(0) = {T;,T2,...,Tp}, where T, > 0 and Tg = T. For t < Ti41, define H;(t) as the time ¢ value of the Bermudan option when exercise is restricted to the dates D(Ti+1) = {Ti41, Tiz2,--., Tp}. That is32 1 Introduction to Arbitrage Pricing ‘Theory : H(t) =NEP (V(Tis1)/N(ina)), 4 1,---,B-1. At time T;, H;(T;) can be interpreted as the hold value of the Bermudan option, that is, the value of the Bermudan option if not exercised at time T;. If an optimal exercise policy is followed, clearly we must have at time T; V(T;) = max(U(T,), Hi(Ti)), i=1,...,B, (1.63) such that H,(t) = N(t)E2” (max (U(Ti41), Hisi(Dia)) /N(Pix)), i= 1,0. B=. (1.64) Starting with the terminal condition Hg(T) = 0, (1.64) defines a useful iteration backwards in time for the value V(0) = Ho(0). We shall use this later for the purposes of designing valuation algorithms in Chapter 18, and for computing price sensitivities (deltas) in Chapter 24. We note that the idea behind (1.63) is often known as dynamic program- ming or the Bellman principle. Loosely speaking, we here work “from the back” to price the Bermudan option. As we shall see later (in Chapter 2), this idea is particularly well-suited for numerical methods that proceed backwards in time, such as finite difference methods. 1.10.1 The Markovian Case We now specialize to the Markovian case where U(t) = g(t, x(t)), where g: (0,T] x R” — R is continuous and de(t) = (t,2(t)) dt +o (t,2(t)) aw(t) (1.65) is an n-dimensional Markovian process, where js and o satisfy the regularity conditions of Theorem 1.6.1. The n-dimensional process!® x(t) here defines the state of the exercise value U(t), so we say that x(t) is a state variable process, For concreteness let our numeraire N(t) be the money market account N(t) = B(t) = elt r(ue(u)) due where the short interest rate r : [0,7] x R" > R is here assumed a function of time and the state variable vector z. In (1.65), W(t) is understood to be a d-dimensional Brownian motion in the risk-neutral measure Q. Writing V(t) = V(t,2(t)), we have from (1.61) V(t,2) = sup B9 (ee reur(u) 449 (7, 2(7))] 2(0) = 2). (1.66) 6Note that «(t) is an abstract construct, and does not necessarily coincide with any asset price process.1.10 Options with Early Exercise Rights 33 For dates t € D(0), clearly V(t,x) > g(t,x), with equality holding only when time t exercise is optimal. This leads us to define the concept of an exercise region as X = {(t,c) € D(0) x R": V(t, x) = g(t,x)}. Similarly, we define the complement of 7, C= {(t,x) € [0,7] xR": (t,2) ¢ X}, to be the continuation region, i.e. the region where we wait (either because exercise is not optimal or because it is not allowed, t ¢ D(0)) rather than exercise the option. For Markovian systems, rather than solving the optimization problem (1.66) directly, it is often particularly convenient to invoke the Bellman principle. Extending the ideas presented earlier, let us, somewhat loosely, state the Bellman principle as follows: for any t € D(0), V(t.) = lim max (ott 1), E2 (eo SA rue) uy (44 Asa(t + 4)))) . (1.67) Again, this simply says that the option value at time ¢ is the maximum of the exercise value and the hold value, that is, the present value of continuing to hold on to the option for a small period of time. As we have seen above, for a Bermudan option, (1.67) also holds for finite A (namely up to the next exercise date). The Bellman principle provides us with a link between present (time t) and future (time t + A) option values that we can often exploit in a numerical scheme. For this, however, we need further characterization of V(t,x) in the continuation region. By earlier arguments, we realize that V(t, x)/B(t) must be a Q-martingale on the continuation region. Assuming sufficient smoothness for an application of Ito’s lemma, this leads to a PDE formulation, to hold for (¢,a) € C, IV (t,x) =0, (1.68) where t, ly Se (t, t, & r(t, x r= sun? HLL ott) ).5 Reda ~ Mh) Assume first that our option is of the Bermudan type, and let 7, and Tra be subsequent exercise dates in the exercise schedule. For any function f of time, define f(t+) to be the limits limeyo f(t + €), and assume that V(Ti41—,2) is known for all «. As all values of t € (Tj,Ti41) by definition must be in the continuation region, we can use (1.68) to solve for V(Tj+, x). Applying the Bellman principle (1.67) at time T; then leads to the condition V (T;-, 2) = max (9(T;, 2), V (Ti+, 2).34 1 Introduction to Arbitrage Pricing Theory In PDE parlance, this is a so-called jump condition which is straightforward to incorporate into a numerical solution; see Section 2.7.4 for details. For American-style options, (1.68) continues to apply on C. The Bellman principle here leads to the characterization that TV(t,x) <0, for (t,r) € #, i.e. we exercise when the rate of return from holding the option strictly fails to match r(t,). The American option pricing problem is often conveniently summarized in a variational inequality, to hold on XY UC, Vita) >9(t.2), IV(t,2) <0, (V(t,x) —9(t,2)) IV(t,2) = 0, (1.69) and subject to the boundary condition V(T, x) = g(T, x). The first of these three conditions expresses that the option is always worth at least its exercise value; the second expresses the supermartingale property of V(t, x); and the third implies (after a little thought) that 7V(t,z) = O0onC and JV(t, x) <0 on X. The system (1.69) is discussed more carefully in Duffie [2001], where additional discussion of regularity issues may also be found. 1.10.2 Some General Bounds In many cases of practical interest, solving PDEs and/or variational in- equalities is not computationally feasible. In such situations, we may be interested in at least bounding the value of an option with early exercise rights. Providing a lower bound is straightforward: postulate an exercise policy 7 and compute the price V7 (0) by direct methods. From (1.60), clearly this provides a lower bound v7(0) < V(0). (1.70) The closer the postulated exercise policy 7 is to the optimal exercise policy r*, the tighter this bound will be. We shall later study a number of numerical techniques to generate good exercise strategies for fixed income options with early exercise rights, see Chapter 18. To produce an upper bound, we can rely on duality results established in Rogers [2001], Haugh and Kogan [2004] and Andersen and Broadie [2004]. Let K denote the space of adapted martingales M for which sup, (0,7) E2” |M(r)| < 00. For a martingale M € K, we then write cap 20% (UD ro (xa) U = an (U(r) 7 ero] (xa + MO) i) vio = Pe Qn U(r) i =M(0)+ sup E Go M(n)).1.10 Options with Early Exercise Rights 35 In the second equality, we have relied on the optional sampling theorem, a result that states that the martingale property is satisfied up to a bounded random stopping time, i.e. that E®’(M(r)) = M(0); see Karatzas and Shreve [1997] for details. We now turn the above result into an upper bound by forming a pathwise maximum at all possible future exercise dates D(0): vrs a (mi) ” U(t) U(t) and A(t) > 0. In conclusion, we have arrived at a dual formulation of the option price ¥¢0) = jn (100) + 82" (a (a3 - mi) } . (7) and have demonstrated that the infimum is attained when the martingale M is set equal to the martingale component of the deflated price process36 1 Introduction to Arbitrage Pricing ‘Vheory V(t)/N(t). In practice, we are obviously not privy to V(t)/N(t) (which is a quantity that we are trying to estimate), but we are nevertheless provided with a strategy to make the upper bound (1.71) tight: use a martingale that is “close” to the martingale component of the true deflated option price process. In Chapter 18 we shall demonstrate how to make this strategy operational. 1.10.3 Early Exercise Premia We finish our discussion of options with early exercise rights by listing some known results for puts and calls, including an interesting decomposition of American and Bermudan option prices into the sum of a European option price and an early exercise premium. For convenience, we work in a Markovian setting where the single state variable, denoted S(t), follows one-dimensional GBMD. Specifically, we assume that dS(t)/S(t) = (r — q)dt +a dW (2), (1.73) with W(t) being a one-dimensional Brownian motion in the risk-neutral measure, i.e. the measure induced by the money market account 8(t) =e”. For simplicity we assume that the interest rate r, the dividend yield g, and the volatility o are all constants; the extension to time-dependent parameters is straightforward. Let c(t), Ca(t), and Ca(t) be the time t European, American, and Bermudan prices of the call option with terminal maturity T, conditional on no exercise prior to time t. While obviously c(t) < Cg(t) < Ca(t), in some cases these inequalities are equalities, as the following straightforward lemma shows. Lemma 1.10.2. Suppose that r > 0 and q <0 in (1.73). It is then never optimal to exercise a call option early, and e(t) = Ca(t) = Ca(t). Proof. Notice that, by Jensen’s inequality, ft) =e -PEP ((S(T) — K)*) + . + > en(T-) ((e (S(T) -K) ) 7 (e890 7 ent) . It is therefore clear that ifr >0 and g < 0, then for any value of T — t, e(t) > (S@) - K)*, ive. the European call option price dominates the exercise value. As the hold value of Ainerican and Bermudan options must be at least as large as the European option price, it follows that the option to exercise early is worthless. O1.10 Options with Early Exercise Rights 37 Remark 1.10.3. For the put option, early exercise is never optimal if r < 0 and q > 0. As this situation rarely happens in practice, American put options nearly always trade at a premium to their European counterparts. Lemma 1.10.2 demonstrates the well-known fact that American or Bermu- dan call options on stocks that pay no dividends (q = 0) should never be ed early. On the other hand, if the stock does pay dividends, for an American call option there will, at time t, be a critical value of the stock, Sa(t), at which the value of the stream of dividends paid by the stock will compensate for the cost of accelerating the payment of the strike K. In other words, an American option should be exercised at time t, provided that S(t) > Sa(t). The deterministic curve S(t) is known as the early exercise boundary and marks the boundary between the exercise and continuation regions, ¥ and C. Writing C(t) = Ca(t, S(t)), we formally have Sa(t) = inf {s = Ca(t,$) =(S- K)*} , t 0 and the high contact condition states that the delta equals —1 at the exercise boundary. Establishing the boundary S'i4(t) will virtually always require numerical methods, although asymptotic results are known for t close to T (see for instance Lipton {2001]). One simple result is listed below. Lemma 1.10.6. Assume that r > 0 and q > 0, such that the early exercise boundary exists for the American call option. The exercise boundary just prior to maturity is then lim $4(T —¢) = K max (. r) . 210 q1.10 Options with Early Exercise Rights 39 Proof. An informal proof of Lemma 1.10.6 proceeds as follows. At time T — dt, assume that S(T — dt) > K; otherwise it clearly makes no sense to exercise the option. If we exercise the option, we receive S(T — dt) — K at time T — dt. On the other hand, if we postpone exercise, at time T — dt our hold value is eT CER y (S(T) — K) = S(T — dtje""" — Ker = S(T —dt)- K — S(T —t)qdt + Krdt. Clearly, we should then only exercise if S(T — dt) — K > S(T ~ dt) — K — S(T — dt)qdt + Krat or if S(T —dt)q> Kr. 5B Notice that since clearly S4(T) = K, the call option exercise boundary will have a discontinuity at time T, if g s4(y}(S(Q) — K). In the continuation region, C'4(t, 5) satisfies the PDE (1.47), ie. aCalt, 8) g2CaltrS) | 12 g20°Calt,S) apt Sag + 9S co Inserting this into (1.79) we get, after a few rearrangements, eCate) =rCa(t, S). dCa(t) = rCa(t) dt + 1siyes,(y)(r — 2) S(t) = aw (t) + ls@rsawny {((r — DS) - Cat ) a + oS(t) dw? (t)} =rCalt)at + lyseesncoylr— 991) 220 away + Ieswrsaiy {(rK — ¢S(t)) dt + oS(t) dw? (t)} Setting y(t) = Ca(t)/B(t), it follows from Ito’s lemma that OC a(t dy(t) =e" Lisaycsa(y}(t — 2)3(t) 40 awe) + lisyesrye™ {(rK — gS(t)) dt + S(t) dw? (t)} . Integrating and taking expectations leads to T EP (¥(T)) = u(t) +f OME (1ys(ay>satw)} (TK — 9S(u))) du Applying the definition of y(t) and the fact that y(T) = e~"7 (S(T) — K)+ proves (1.76). The explicit form of the early exercise premium in (1.78) follows from the properties of GBMD. 0 Remark 1.10.8. Combining results from Lemma 1.10.6 and Proposition 1.10.4, it follows that E(t) > 0, so Ca(t) > c(t) as expected. The integral representation of the American call option in Proposition 1.10.7 forms the basis for a number of proposed computational methods for American option pricing. Loosely speaking, these methods are based on the idea of iteratively estimating the exercise boundary S(t), often working backwards from ¢ = T, after which an application of Proposition 1.10.7 will yield the American option price. A representative example of these methods can be found in Ju [1998]. See Chiarella et al. [2004] for a survey of the literature, and Section 19.7.3 for an application in interest rate derivative pricing. For a Bermudan option, an integral representation such as that in Propo- sition 1.10.7 is not possible. Nevertheless, it is still possible to break the1.10 Options with Early Exercise Rights 41 Bermudan call option into the sum of a European option and an early exercise premium. To show this, assume that the allowed exercise dates are D(0) = {T1,T2,...,T pz}, and let Sg(T;) be the exercise level above which the Bermudan option should be exercised at time T;, i = 1,...,B. Notice that if at time J; we have S(T;) > Sg(T;), then Cg will jump down in value when time progresses past time Tj, as a reflection of the missed exercise opportunity. Indeed, in the earlier notation of hold and exercise values, we have Cp(Ti) = max (U(T;), H(Ti)), Cn(Ti+) = H(T), which makes the jump in value evident. Given the existence of these jumps, we may write dCp(t) = rCp(t) dt + dM(t) B + > Lps¢r,)>sp(72)}9(Ti — t) (H (Ti) — U(T)) dt, i=l where H(T;) = Cg(T;+) is the hold value at time T;, U(T;) = S(T;) — K, and M(t) is a martingale, OCz(t) as" Deflating Cg by the money market account and forming expectations, we get, since c(t) = e779 E2(CR(T)), Cal) = ct) + Soe PRP (1p scr) >sacn} (UT) ~ H(Ti))) - Ti>t dM(t) = a) S(t) dw? (2). As H(T;) must be less than the exercise value U(T;) whenever $(T;) > Sp(Ti) we can simplify this expression to the following result. that we, in Section 18.2.3, call the marginal exercise value decomposition. Proposition 1.10.9. The Bermudan option price Cg(t) satisfies Ca(t) =clt)+En(t), t R are sufficiently regular to make (2.1) and (2.2) meaningful (see Chapter 1). The terminal value problem above is, as discussed earlier, a Cauchy problem to be solved for V(t,x) on (t,x) € {0,T) x B. In many cases of practical interest, further boundary conditions are applied in the spatial (x) domain. If such boundary conditions are expressed directly in terms of V (rather than its derivatives) we have a Dirichlet boundary problem. For instance, a so-called up-and-out barrier option will pay out g(x(T)) at time T if and only if x(t) stays strictly below a contractually specified barrier level H at all times t < T. If, on the other hand, x(t) touches H at any time during the life of the contract, it will expire worthless (or “knock out”). In this case, the PDE is only to be solved on (t,x) € [0,T) x (BN (—00, H)) and is subject to the Dirichlet boundary condition V(t,H) =0, t€ (0,7), which expresses that the option has no value for z > H. We note that it is not uncommon to encounter options where the spatial domain boundaries are functions of time, a situation we shall deal with in Section 2.7.1. Also, as we shall see shortly, sometimes boundary conditions are conveniently expressed in terms of derivatives of V. For numerical solution of the PDE (2.1), we often need to assume that the domain of the state variable z is finite, even in situations where (2.1) is supposed to hold for an infinite domain. Suitable truncation of the domain can often be done probabilistically, based on a confidence interval for x(T). To illustrate the procedure, consider the Black-Scholes PDE (1.47) applied to a call option with strike K. A common first step is to use the transformation zx = In S, such that the PDE has constant coefficients, av 15\8 1 40°V _ a+ 0) 4 3 rv =0, (2.3)2.2 Finite Difference Discretization 45 with terminal value (for a call option) V(T, x) = (e*—K)*. The domain of z is here the entire real line, B = R. We know (from (1.39)) that 2(T) = 2(0) + rte? T +0(W(T)-W(0)), (2.4) 2 which is a Gaussian random variable with mean % = x(0) + (r — 40)T and variance o?T. Consider now replacing the domain (—oo, 00) with the finite interval [= — ao VT, + ac VT] for some positive constant a. The likelihood of z(L) falling outside of this interval is easily seen to be 26(—a) (where, as always, ®(z) is the standard Gaussian cumulative distribution function). If, say, we set a to 4, 26(—a) = 6.3 x 1075, which is an insignificant probability for most applications. Larger (smaller) values of a will make the truncation error smaller (larger) and will ultimately require more (less) effort in a numerical scheme. We recommend values of a somewhere between 3 and 5 for most applications. For the Black-Scholes case, a rigorous estimate of the error imposed by domain truncation is given in Kangro and Nicolaides (2000). In many cases of practical interest, it is not possible to write down an exact confidence interval for «(T). In such cases, one instead may use an approximate confidence interval, found by, for instance, using “average” values for (t,x) and o(t, x). High precision in these estimates is typically not needed. 2.2 Finite Difference Discretization In order to solve the PDE (2.1) numerically, we now wish to discretize it on the rectangular domain (t,.:) € [0,7] x [M, M], where M7 and M are finite constants, possibly found by a truncation procedure such as the one outlined above. We first introduce two equidistant! grids {t;}.9 and (as}jco where t, =iT/n £ iA, i =0,1,...,n, anda; = M+j(M-M)/(m+1) = M+jA,, j=0,1,...,m+1. The terminal value V(T, x) = g(x) is imposed at tn = T, and spatial boundary conditions are imposed at t and £41. 2.2.1 Discretization in x-Direction. Dirichlet Boundary Conditions We first focus on the spatial operator £ and restrict x to take values in the interior of the spatial grid x € {x;}7%,. Consider replacing the first- and second-order partial derivatives with first- and second-order difference operators: 1Non-equidistant grids are often required in practice and will be covered in Section 2.4.46 2 Finite Difference Methods V(t, aj41) — V(t, 27-1) 5,V(t, xj) © Cy nn (2.5) Vit. Vi ~1) — 2V(E, SnaV (t,0;) & Ve tesa) + V(b 25-1) = V(b 25) es 1) = 2V (tj) (2.6) 3 These operators are accurate to second order. Formally”, Lemma 2.2.1. OV (t,x 5,V(t,25) = Wiis) +0(42), PV (ta SneV (t, 05) = oVieee) +0(42). Proof. A Taylor expansion of V(t, x) around the point x = x; gives OV (t, 25 V(tuatj41) = V(t.2,) + Oe PV(t,a;) , 1.3 BV (t,25) +5 5A + Eat aa +0 (42), and OV(t, 25 V(tyay-1) = V(t) — 1 OV(t,2;) 1 ,30°V(t, 23) 4 +5ap Mea) _ 7 ag PVE) 5 o (a8). Insertion of these expressions into (2.5) and (2.6) gives the desired result. a In other words, if we introduce the discrete operator L = ult,2)b_ + 50(t,2)*6ex —r(t,2), we have, for x € {x;}71, LV (t,x) = LV(t,2) +O (A2). With attention restricted to values on the grid {2;}72,, we can view Lasa matrix, once we specify the side boundary conditions at xp and 241. For the Dirichlet case, assume for instance that V(20,t) =F (t,%0), V(amsrst) =F (t,m41), ?Recall that a function f(h) is of order O(e(h)) if [f(h)|/le(h)| is bounded from above by a positive constant in the limit h > 0.2.2 Finite Difference Discretization AT for given functions f,f : (0,7) x R — R. With? w(t) 4 (V(t,21),---,V(t,em))" and, for j = 1, , o(t) & —o(t,ay)?AZ? — r(t, x5), (2.7) 1 1 ug(t) © sm(t,aj)Az? + got ei) AS*, (2.8) Ut) & —Sultsay)g" + Solty2y)?Az?, (2.9) we can write 7 LV(t) = A(t) V(t) + QU), (2.10) where A is a tri-diagonal matrix ex(t) u(t) 0 0 0 0 l(t) co(t) w(t) 0 0 0 0 I(t) s(t) us(t) 0 0 A(ty=| 2 9 ut ca(t) ua(t) 0 (2.11) : 7 : . 7 . 0 000. bmualt) nes) taf) 0 0 0 0 0 Im(t) em(t) and Q(t) is a vector containing boundary values L(t)F (t,o) 0 X= : _0 Um (t)F (t, mst) As discussed earlier, sometimes one or both of the functions f and f are explicitly imposed as part of the option specification (as is the case for a knock-out options). In other cases, asymptotics may be necessary to establish these functions. For instance, for the case of a simple call option on a stock paying no dividends, we can set F(t,2) =e — Ket, f(t,x) =0, where we, as before, have set « = In S (S being the stock price) and assumed that the strike K is positive. The result for f is obvious; the result for F follows from the fact that a deep in-the-money call option will almost certainly pay at maturity the stock (the present value of which is just S = e*) minus the strike (the present value of which is Ke7"(?-)). 3For clarity, this chapter uses boldface type for all vectors and matrices.48 2 Finite Difference Methods 2.2.2 Other Boundary Conditions Deriving asymptotic Dirichlet conditions can be quite involved for compli- cated option payouts and is often inconvenient in implementations. Rather than having to perform an asymptotic analysis for each and every type of option payout, it would be preferable to have a general-purpose mechanism for specifying the boundary condition. One common idea involves making assumptions on the form of the functional dependency between V and x at the grid boundaries, often from specification of relationships between spatial derivatives. For instance, if we impose the condition that the second derivative of V is zero at the upper boundary (2.41) — that is, V is a linear function of « — we can write (effectively using a downward discretization of the second derivative) V(t, ¢m41) + V(t,em—1) — 2V(t, 2m) _ 0 a2 ~ ie => V(t, tm+i) = 2V(t,2m) — V(t, em~1)- A similar assumption at the lower spatial boundary yields V(t, 20) = 2V(t,21) — V(t»). For PDEs discretized in the logarithm of some asset, it may be more natural to assume that V(t,x) « e* at the boundaries; equivalently, we can assume that OV/Ox = 0°V/dx? at the boundary. When discretized in downward fashion at the upper boundary (241), this implies that. V(tytmar) —V (ttm) _ V(t,tme1) + V(t, ¢imai) = 2V (tem) Ay a or (assuming that A, # 1) 1 V(t,m41) = V(t tm—1) A =a+ V(t tm) A xT Similarly, a 2+ Ar _ . 1 V(t, 20) = V(t.) 7 +a, Vibe)as Common for both methods above — and for the Dirichlet specification discussed earlier — is that they give rise to boundary specifications through simple linear systems of the general form V(t, @m41) = kmn(t)V (tim) + km—1(t)V (ts @m—1) + f (t)@m41), (2.12) V(t, 20) = ki(t)V(t, v1) + ko(t)V(t, a2) + f(t, 20). (2.13) This boundary specification can be captured in the matrix system (2.10) by simply rewriting a few components of A(t); specifically, we must set2.2 Finite Difference Discretization 49 * = 1(t, 0m) + Fm(t)Um(t)s 1 1 Im(t) = =5u(t,tm)As* + 50(t, tm) Az? + hm—a(t}um(t), ex(t) = -o(t,2;)?AZ? — r(t,25) + k(t), em(t) = —9(t, tm)” 1 a(t) = Zultsor)Ag! + 5o(te)?Ag? + halt lt. All other components of A remain as in (2.11); note that A remains tri- diagonal. An alternative approach to specification of boundary conditions in the a-domain involves using the PDE itself to determine the boundary condi- tions, through replacement of all central difference operators with one-sided differences at the boundaries. Section [Link] contains a detailed example of this idea; ultimately, this approach leads to boundary conditions that can also be written in the form (2.12)-(2.13). 2.2.3 Time-Discretization To simplify notation, assume for now that Q(t) = 0 for all t, as will be the case if, say, we use the linear or linear-exponential boundary conditions outlined earlier. On the spatial grid, our original PDE can be written “sO = —A(t)V(t) + O (A?) which, ignoring the error term*, defines a system of coupled ordinary differ- ential equations (ODEs). A number of methods are available for the numerical solution of coupled ODEs; see, e.g., Press et al. [1992]. We here only consider basic two-level time-stepping schemes, where grid computations at time ¢; involve only PDE values at times t; and t;41. Focusing the attention on a particular bucket [t;,t;41], the choice for the finite difference approximation of OV/At is obvious: OV _ Vitiss)- Viti) ot At . Not so obvious, however, is to which time in the interval [t;,¢;41] we should associate this derivative. To be general, consider picking a time t}+1(0) € (ti, tis], given by 41 (6) = (1 — O)tin + ti, (2.14) where € (0, 1} is a parameter. We then write BV (6) | Vitus) = Vite) at A, 7 ‘Note that the error term O(A2) is here to be interpreted as an m-dimensional vector. We will use such short-hand notation throughout this chapter.50 2 Finite Difference Methods By a Taylor expansion, it is easy to see that this expression is first-order accurate in the time step when 6 # 3, and second-order accurate when 6 = }. Written compactly, av (18 tigi) — V(ti we") = We ts +1 fo4}0 (4) +0 (42). (2.15) This result on the convergence order is intuitive since only in the case 9 = 2 is the difference coefficient precisely central; for all other cases, the tite ence coefficient is either predominantly backward i in time or predominantly forward in time. The time-discretization technique introduced above i is known as a theta scheme. The special cases of @ = 1, 6 = 0, and 6 = 4 are known as the fully implicit scheme, the fully explicit scheme, and the Crank-Nicolson scheme, respectively. In light of the convergence result (2.15), one may wonder why anything other than the Crank-Nicolson scheme is ever used. The CN method is, indeed, often the method of choice, but there are situations where a straight application of the Crank-Nicolson scheme can lead to oscillations in the numerical solution or its spatial derivatives. Judicial application of the fully implicit method can often alleviate these problems, as we shall discuss later. The fully explicit method should never be used due to poor convergence and stability properties (see Section 2.3), but has nevertheless managed to survive in a surprisingly large number of finance texts and papers. 2.2.4 Finite Difference Scheme We now proceed to combine the discretizations (2.10) and (2.15) into a complete finite difference scheme. First, we expand A (HB) V (ti°2(0)) = 8A (117(8)) Viti) + (1-8) A(47*(6)) V(tier) + Lgoey} (Ae) + 0 (42), such that our PDE can be represented as Vitis) — Viti) ——A + 149 44}0(A1) + O (A?) =A (G8) V (EP @) +0 (42) = — OA (t)**(8)) V(t.) — (1 = 0) A (47*(6)) V(tins) + Logg} (Ar) + O (4?) +0 (A2). Multiplying through with A, gives rise to the complete finite difference representation of the PDE solution at times t; and t;41:2.2 Finite Difference Discretization 51 Proposition 2.2.2. On the grid {xj}, the solution to (2.1) at times ty and ti41 is characterized by (I-6A,A (t)*1(8))) V(ti) = (I+ (1 - @) ALA (#177(0))) V(tign) + ei, (2.16) ith where I is the m x m identity matrix, and e;** is an error term 1 = AcO (AZ) + 1f943}0 (Az) +0 (42). (2.17) Let V(ti,2) denote the approximation to the true solution V(t), obtained by using (2.16) without the error term. Defining Vi) = (Pea)... Pam), we have (1-04, A (t**1(0))) V(t,) = (T+ (1 — 8) ALA (t°47(0))) V(tign). (2.18) For a known value of Vitins)s (2.18) defines a simple linear system of equa- tions that can be solved for Viti) by standard methods. Simplifying matters is the fact that the matrix (I-9A,A(t{*?(0))) is tri-diagonal, allowing us to solve (2.18) in only O(m) operations; see Press et al. [1992] for an algorithm®. Starting from the prescribed terminal condition V(tn,2)) = g(x), J = 1,...,m, we can now use (2.18) to iteratively step backward in time until we ultimately recover V(0). This procedure is known as backward induction. Proposition 2.2.3. The theta scheme (2.18) recovers V(0) in O(mn) op- erations. If the scheme converges, the error on V(0) compared to the exact solution V(0) is of order 0 (A2) + 1¢543}0 (Ar) +0 (42). Proof. The backward induction algorithm requires the solution of n tri- diagonal systems, one per time step, for a total computational cost of O(mn). The local truncation error on V(t;) is e't?, making the global truncation error after n time steps of order ne'+!. Combining (2.17) with the fact: that n= T/A; = O(A;') gives the order result listed in the proposition. O The special case of an explicit scheme (@ = 0) provides us with a direct expression for V(t,,2;) in terms of V(ty41,2j-1), V(lig1, 2), and V(ti21, 2541), a scheme that is easily visualized as a “trinomial tree”. The intuitive nature of the explicit scheme coupled with the fact that no matrix equation must be solved may explain the popularity of this scheme in the finance literature, despite its poor numerical qualities (see Section 2.3). We stress that the workload of the explicit scheme is still O(m) per time step, as is the case for all theta schemes.52 2 Finite Difference Methods It follows from Proposition 2.2.3 that the Crank-Nicolson scheme is second-order convergent in the time step, and all other theta schemes are first-order convergent in the time step. All theta-schemes are second-order convergent in the spatial step Az. In deriving (2.18), we assumed earlier that the boundary vector was zero, Q(t) = 0. Including a non-zero boundary vector into the scheme is, however, straightforward and results in a time-stepping scheme of the form (I-0A,A (t#*1())) V(t) = (I+ (1 — 8) ALA (t3*1(0))) V (tig) + (1~ 0)2%ti41) + OME). (2.19) Again, this system is easily solved for Viti) by a standard tri-diagonal equation solver. As a final point, we stress that the finite difference scheme above ulti- mately yields a full vector of values Vo) at time 0, with one element per value of z;, j =1,...,m. In general, we are mainly interested in V(0, x(0)), where «(0) is the known value of x at time 0. There is no need to include x(0) in the grid, as we can simply employ an interpolator (e.g., a cubic spline) on this vector V(0) to compute V(0, x(0)). Clearly, such an interpolator should be at least second-order accurate to avoid interfering with the overall O(A?) convergence of the finite difference scheme. Assuming the interpolator is sufficiently smooth, we can also use it to compute various partial derivatives with respect to x that we may be interested in. Alternatively, these can be computed by the same type of finite difference coefficients discussed in Section 2.2.1. The derivative OV(0,x(0))/Ot — the time decay — can be picked up from the grid in the same fashion. Remark 2.2.4. The scheme (2.18) may, without affecting convergence order, be replaced with (I-OA,A(t:)) V(ti) = (+ 0 - A) ALA (t41)) Vig). 2.3 Stability 2.3.1 Matrix Methods Ignoring the contributions from boundary conditions, the finite difference scheme developed in the previous section can be rewritten V(t) = BEV (tay), (2.20) where Byt? S (1-04, A (ti*7(0)))~ * (E+ (1-8) ALA (ti*4(0))).2.3 Stability 53 That is, for any 0S k Miltaje i, L where H;(t;,) and w are the amplification factor (discrete Fourier transform) and wave number for the /-th mode, respectively. Notice that i here denotes ©The spectral norm of a matrix C is defined as the largest absolute eigenvalue of (C’C)'/?. ‘Phe infinity norm is defined as max; }); |Ci,y|- 7In the application to PDEs with non-constant coefficients, it may help to think of the von Neumann analysis as being applied to the PDE locally with “frozen” coefficients, followed by an examination of the worst case among all frozen coefficients.54 2 Finite Difference Methods the imaginary unit, i? = —1, with k (momentarily) having taken the role of the time index in the finite difference grid. For the constant coefficient case, a key fact for our PDE problem is that Hi(te) = Hiltes)&r"s where & is a mode-specific amplification factor independent of time. To determine how a solution is propagated back through the finite difference grid, it thus suffices to consider a test function of the form (ty sj) = E(w) Rete (2.22) According to the Von Neumann criterion, stability of (2.20) requires that the modulus of the amplification factor E(w) is less or equal to one, independent of the wave number: Vw : |E(w)| <1. (2.23) This criterion is natural and merely expresses that all eigenmodes should be dampened, and not exponentially amplified, by the finite difference scheme. Turning to our system (2.20), assume for simplicity that r(t,«) =0. A positive interest rate (we will nearly always have r(t,%) > 0) introduces some extra dampening through discounting effects and will, if anything, lead to better stability properties than the case of zero interest rates. Writing v([Link]) = veg, O(tET (8), ay) = oe, and w(t{**(9),23) = jue, the von Neumann analysis gives the following result: Proposition 2.3.1. Define a = A,/(Az)*. For (2.20) with r(t,x) = 0, the von Neumann stability criterion is 2 Chg ; (2.24) mle Nok g + hg At + [125 Az = 08.5 to hold for all k =0,1,...,n—1, j= 1,2,....m. Proof. Define sj, = o7,; + Arus,j- A local application of (2.20) gives ad _ ad Uk j-1 (-Sa,) + vg (1 + aO0% 5) + Ue 544 (-¥st,) = a(1—@) _ 3 1-@ deans (7s) +ups1(l—a(L- op) Hee et (7) with a defined above. Inserting (2.22) and rearranging (using Euler’s formulas for sin and cos) yields 1— (1 —A)aoZ (1 — coswAg) + i(1 — AoAg py, SinwAy 1+ dao ;(1 — coswAz) — Bas j1x,j 8inwAy . Ew) =2.3 Stability 55 Note that € is a function of k and j, due to the non-constant PDE parameters. As discussed earlier (see also Mitchell and Griffiths [1980]), we expect. the system to be stable if the criterion (2.23) holds for all k and j in the grid. Computing the modulus of € and requiring that it does not exceed one leads, after straightforward manipulations, to the stability criterion Ve: 2aog,; + (26 1)0? fof, + ui AZ + coswAs (uz gAz — o%,;)] 2 0. As coswA, € [-1, 1], this expression can be simplified to (2.24). 0 From (2.24) we can immediately conclude that the finite difference scheme is always stable if 3 <@ <1, irrespective of the magnitudes of A, and A;. For 3 < @ <1, we therefore say that the theta scheme is absolutely stable, or simply A-stable. Both the fully implicit (@ = 1) and the Crank-Nicolson (6 = 4) finite difference schemes are thus A-stable. For the explicit scheme (@ = 0), however, stability is conditional, requiring 2 2 4 2 2 2 2 4 Fok g 2 Thy + MRA + [Meg AE — oh | For small drifts, this expression amounts to the restriction oR, < a2 /A, which can be quite onerous, often requiring the (laborious) use of thousands of time steps in the finite difference grid. We shall not consider fully explicit methods any further in this book. Returning to the case } < @ <1, let us introduce a stronger definition of stability. A time-stepping method is said to be strongly A-stable if the modulus of the amplification factor € is strictly below 1 for any value of the time step, including the limit® A, + oo. From (2.24), we see that if A, + 00 (which implies a — 00), then the modulus of the amplification factor could reach 1 in the special case of @ = 1/2. In other words, the Crank-Nicolson scheme is not strongly A-stable. For large time steps, harmonics in the Crank- Nicolson finite difference solution will effectively not be dampened from one time step to the next, opening up the possibility that unwanted high- frequency oscillations can creep into the numerical solution. In practice, this is primarily a problem if high-frequency eigenmodes have high amplification factors, as can happen if there is an outright discontinuity in the terminal value function g. The problem is especially noticeable if the discontinuity in the value function is “close” in both time and space to t = 0 and x = (0) (as would be the case for a short-dated option with a discontinuity close to the starting value of x). Oscillations can be prevented by setting the time step smaller than twice the maximum stable explicit time step (see Tavella and Randall [2000]), but this can often be computationally expensive. We shall deal with other methods to suppress oscillations in Section 2.5. We conclude this section by noting a deep connection between the stability of a finite difference scheme and its convergence to the true solution if further |é| approaches zero for At —> 0, the scheme is said to be L-stable.56 2 Finite Difference Methods of the PDE as A; > 0 and A, — 0. First, we define a finite difference scheme to be consistent if local (Taylor) truncation errors approach zero for A, + 0 and A, — 0. All the schemes we have encountered so far are consistent. Further, define a finite difference scheme to be convergent if the difference between the numerical solution and the exact PDE solution at a fixed point in the domain converges to zero uniformly as A; > 0 and A, —> 0 (not necessarily independently of each other). We then have Theorem 2.3.2 (Lax Equivalence Theorem). For a well-posed® linear terminal value PDE, a consistent 2-level finite difference scheme is convergent if and only if it is stable. A more precise statement of the above result, as well as a proof, can be found in Mitchell and Griffiths [1980]. 2.4 Non-Equidistant Discretization In practice, we often wish to align the finite difference grid to particular dates (e.g., those on which a coupon or a dividend is paid) and particular values of x (e.g., those on which strikes and barriers are positioned). Also, for numerical reasons we may want to make certain important parts of the finite difference grid more densely spaced to concentrate computational effort on domains of particular importance to the solution of the PDE. To do so, we will now relax our earlier assumption of equidistant discretization in time and space. Doing so for the time domain is actually trivial and merely requires us to replace A; in (2.18) with Ay; © ti41 — ti, where the spacing of the time grid {t;}/-p is now no longer constant. The backward induction algorithm can proceed as before. We note that the ability to freely select the time grid will allow us to line up perfectly with dates that carry high significance for the product in question (e.g. dates on which cash flows take place, see Section 2.7.3) or to, say, use coarser time steps for the part of the finite difference grid that is far in the future. For an adaptive algorithm to automatically select the time-step, see d’Halluin et: al. [2001]. For the spatial step, we have a number of options to induce non- equidistant spacing. One method involves a non-linear change of variables y = h(x) in the PDE, followed by a regular equidistant discretization in the new variable y. This maps into a non-equidistant discretization in x which, provided that h(-) is chosen carefully, will have the desired geometry. Discussion of this method along with guidelines for choosing h(-) can be found in Chapter 5 of Tavella and Randall [2000]. We will here pursue a more direct alternative, where we simply introduce an irregular grid {aj }7™" *Well-posed means that the PDE we are solving has a unique solution that depends continuously on the problem data (PDE coefficients, domain, boundary conditions, etc.)2.4 Non-Equidistant Discretization 57 and redefine the finite difference operators (2.5)~(2.6) to achieve maximum precision. For this, define + a ~ = Ar; = tj41- 23, AL; = Uj—1s and set _ Vtej41) — Vas) t,2j) — 6fV (taj) = ar (ty 25 ro Vv, » OV (te) = By a Taylor expansion, we get OV (t,x 1&V(t, 82V 623) = WE), SO at, 1PV(t.e 2 pein) 2) (43,)"+0((45,)"), (2.25) _ OV(t,2;) 10?V(t,23) ._ 45 V bas) = Oe aa, 10°V(t,2;) ,,— \2 ~ 3 + gar (425) +0((43,)°). (2.26) Maximum accuracy on the first-order derivative approximation is achieved by selecting a weighted combination of (2.25)-(2.26) such that the terms of order O(A?,) and O(A; ,) cancel. That is, we set 52V (t, 23) = Ae sty 15) A acvee xj) (2.27) V(t, 25) = eV (td V(t, 9 . At, +45; c a 7 x ‘J 24+ — \2 — W(t 05) |g ( (Ady) ey + (Aes) 455 Ox At; +45; which is second-order accurate, in the sense that reducing both A}; and Az, by a factor of k will reduce the error by a factor of k?. To estimate the derivative 0°V(t,2r,)/Ox? we set 5¢V (t,x) - Jz V(t,25) SeeV (t, 2j) = 2 nai 3 (AE; + 425) 2 — \2 3 — \3 _ PV (23) , 9 ( (Az) — (Aey) , (Azs) + (Asy) Ox? At, +47; At ,+ Az; (2.28) which is only first-order accurate, unless At, = Az. Despite this, the global discretization error will typically remain second-order in the spatial step, even for a non-equidistant grid. A proof of this perhaps somewhat58 2 Finite Difference Methods surprising result can be found in the monograph Axelsson and Barker [1991] on finite element methods. Development of a theta scheme around the definitions (2.27) and (2.28) proceeds in the same way as in Section 2.2. The resulting time-stepping scheme is identical to (2.18), after a modification of the matrix A. Specifically, we must simply redefine the c-, u-, and l-arrays in (2.7)-(2.9) as follows: = olt, 25)" — v(t, 23), (2.29) A; 42; 7 A. uj(t) © q(t, 05) + alt, 5), (2.30) (45; +475) 42; (43+ 47; ) Az; At 1 L(t) £4 -————) ry 2 0 * aay ae) + GE aay an (2.31) For an example where having a non-equidistant grid is essential to the numerical performance of the scheme, see Section 9.4.3. 2.5 Smoothing and Continuity Correction 2.5.1 Crank-Nicolson Oscillation Remedies As discussed earlier, for discontinuous terminal conditions, the Crank- Nicolson scheme may exhibit localized oscillations if the time step is too coarse relative to the spatial step. Depending on the timing and spatial position of the discontinuities, these spurious oscillations may negatively affect the computed option value or, more likely, its first (“delta”) or second (“gamma”) x-derivatives. Further, in the presence of discontinuous terminal conditions, the expected O(A?) convergence order of the Crank-Nicolson scheme may not be realized. While O(A?) convergence is possible without spurious oscillations in some multi-level time-stepping schemes, there is evidence that these schemes are less robust than the Crank-Nicolson scheme for many financially relevant problems, see, e.g., Windcliff et al. [2001]. For- tunately, it is relatively easy to remedy the problems in the Crank-Nicolson scheme. Specifically, a theoretical result by Rannacher [1984] shows that second-order convergence can be achieved for the Crank-Nicolson scheme, provided that two simple algorithm modifications are taken: © The discontinuous terminal payout is least-squares (L?) projected onto the space of linear Lagrange basis functions!°. 1Recall that the linear Lagrange basis functions (also called “hat” functions) are simply small triangles given by lj(x) = l{z,_1<2<2j}* a jo +12, H}, where the level H (the digital strike) is located between nodes x, and60 2 Finite Difference Methods x41. For nodes 2;, 7 > k+ 1, clearly V(T,2;) = 1; for nodes x;, j < k, V(T, xj) = 0. The smoothing algorithm will have effect only at xg or Ze41, and will set either V(T,z,) or V(T, 2441) to a value somewhere between 0 and 1, depending on which of xg or xg41 is closest to H. If H happens to be exactly midway between a, or £441, the continuity correction is seen to have no effect whatsoever. The digital option example above gives rise to a method listed in Tavella and Randall [2000] (see also Cheuk and Vorst {1996]). Here, we simply arrange the spatial grid such that the x-values where the payoff (or its derivatives) is discontinuous are exactly midway between grid nodes. If necessary, we can use a scheme with non-equidistant grid spacing to accomplish this (see Section 2.4). Our example above shows that aligning the grid in this way will, in a loose sense, make the payoff smooth. For digital options, the grid shifting technique can be very efficient, and such “locking” of the location of strikes and barriers relative to the spatial grid can often reduce odd-even effects even better than the continuity correction discussed earlier. To demonstrate, consider the concrete task of using a finite difference grid to price a digital call option on a stock S$ in the Black-Scholes model. In this case, we conveniently have a theoretical option price to compare against, since it is easily shown that the time 0 value V(0) must be vo) = In(S(0)/H) + (r 0/2) “TQ (S(T) > H)=e77® ( (2.33) For our numerical work, we discretize the asset equidistantly in log-space (i.e., we work with the PDE (2.3)) and determine the spatial grid boundaries by probabilistic means using a multiplier of a = 4.5, see Section 2.1. Spatial boundary conditions are 9V/Ox = 0?V/Ax?, implemented as described in Section 2.2.2. In one experiment, we apply a straight Crank-Nicolson approach, with no attempt to regularize the payoff condition. In a second experiment, we combine Crank-Nicolson with Rannacher stepping and also nudge the entire spatial grid upwards until the log-barrier In(H) is located exactly half-way between two spatial grid points. Numerical results are shown in Figure 2.1. As Figure 2.1 shows, a naive Crank-Nicolson implementation is plagued by severe odd-even effects and very slow convergence — 100’s of spatial steps appear to be necessary before acceptable levels of the option price are reached. On the other hand, grid shifting combined with Rannacher stepping results in a perfectly smooth!! convergence profile, and 5-digit price precision is here reached in less than 30 steps. 11 can be verified that the convergence order in m is, as expected, close to 2 in this experiment.2.6 Convection-Dominated PDEs 61 Fig. 2.1. 3 Year Digital Option Price — Grid Shifting 0.30 —* Straight Crank-Nicolson 0.25 + + 5 1S 25 35 45, Grid Points in Asset Direction (m ) Notes: Finite difference estimates for the Black-Scholes price of a 3 year digital option with a strike of H = 100. The initial asset price is S(0) = 100, the interest rate is r = 0, and the volatility is o = 20%. ‘lime stepping is performed with an equidistant grid containing n = 50 points. Spatial discretization in log-space is equidistant, as described in the main text; the number of grid points (m) is as listed on the a-axis of the figure. The “Straight Crank-Nicolson” graph shows the convergence profile for a pure Crank-Nicolson finite difference grid. ‘The “Grid Shifting” graph shows the convergence profile for a Crank-Nicolson finite difference grid with Rannacher stepping and a shift of the spatial grid to center In(H) midway between two grid points. From (2.33), the theoretical value of the option is 0.4312451. 2.6 Convection-Dominated PDEs Recall from Section 2.3 that stability of the explicit finite difference scheme requires that (omitting grid subscripts on y and o) 247 2 A 2 Ad 2 A2 4 a, ° Dott pw? + [pA? - 04. As discussed, this condition can be violated if A; is too large relative to A,. However, for fixed A; and A, we notice that instability can also be triggered if the absolute value of the drift jz is raised to be sufficiently large relative to the diffusion coefficient . While theta schemes with @ > 1/2 are always stable, large drifts in the PDE can nevertheless cause spurious oscillations and an overall deterioration in numerical performance of these schemes. PDEs for which this effect62 2 Finite Difference Methods occurs are said to be convection-dominated. To quantify matters, assume for simplicity that the finite difference grid is equidistant in the x-direction, and consider the matrix A in (2.11) with tri-diagonal coefficients c, u, and 1 given by (2.7)-(2.9). As discussed in e.g. d’Halluin et al. [2005], spurious oscillations can occur when, for some t and some j, either uj(t) < 0 or 1;(t) < 0. From (2.8) and (2.9), to avoid spurious oscillations we would thus need alts)? > |u(t,«;)[Ae. (2.34) Intuitively, in convection-dominated systems, the central difference coefficient 6, and d;¢ used to discretize the PDE can no longer fully contain the large expected up- or downward trend of the underlying process for 2; as a result, spurious oscillations can occur. 2.6.1 Upwinding There are a number of well-established techniques to deal with convection- dominated PDEs. First, we can obviously attempt to lower A, such that (2.34) is satisfied. This, however, may not be practical from a computational standpoint (and may require that A; is lowered as well to avoid spurious oscillations originating from the time-stepping scheme). An alternative is to modify the first-order discrete operator 6, such that it points in the direction of the large absolute drift. For instance, we can simply elect to use a suitably oriented one-sided difference, rather than a central difference, whenever (2.34) is violated. This procedure is known as upstream differencing or upwinding. To formalize the idea, introduce a new first-order difference operator 5% given as 3 (V(t, 541) —V(tyaj-1)) Az, [a(t ary) Ae S a(t, 25)?, 52V(t, 23) = 4 (V(t, 23) — V(t, 23-1) Az!, w(t, 2;)A, < —o(t, x;)?, (V(t, c541) — V(t,2j)) Apt, u(t, x;)Ag > a(t, x3). Using 65% instead of 5, modifies the matrix A in (2.11). Specifically, if u(t,xj)A, < —a(t, xj)? we replace (2.7)-(2.9) with: ej(t) = w(t, 2s) AZ! - oft,25)?Az? —r(t, a5), (2.35) 1 u;(t) = 50 (tay) Az? (2.36) -1,1 2 A-2 y(t) = —wlt,25) Ag! + 5o(t,25)7Az?. (2.37) And when p(t,2;)Az > o(t,x;)?, we use oj(t) = —n(t, xj)! — o(t, 3)? Az? — r(t, a5), (2.38) -1,1 - uj(t) = y(t, 23)AZ* + 37a) Az”, (2.39) L(t) = sat )2Az?, (2.40)2.7 Option Examples 63 For non-equidistant grids, a similar modification to (2.29)-(2.31) is required. We omit the straightforward details. Let us try to gain some further understanding of the upwind algorithm. Comparison of (2.35)-(2.40) with (2.7)-(2.9), shows that upwinding amounts to using a regular central difference operator 6, on a PDE with a diffusion coefficient modified to be o(t,z) + /|u(t,x)|Ax. The numerical scheme in effect introduces enough artificial diffusion into the PDE to satisfy (2.34). Doing so, however, comes at a cost: the convergence order of the scheme will be reduced to O(A,) if one-sided differencing ends up being activated in a significant part of the grid. We note that higher-order upwinding schemes are possible if the finite difference operator 5% is allowed to act on more than three neighboring points. For such schemes, the matrix A will no longer be tri-diagonal. 2.6.2 Other Techniques As discussed earlier, upwinding amounts to adding numerical diffusion at nodes where the scheme is convection dominated. Alternatively, we can increase a(t, x) directly, to o(t,x) + € where ¢ is chosen to be large enough for the scheme to satisfy (2.34). By solving the resulting PDE for different values of ¢, it may be possible to determine how the error associated with € scales in e. This, in turn, will allow us to extrapolate to the limit ¢ = 0. See p. 135 of Tavella and Randall [2000] for an example. The upwinding scheme presented in Section 2.6.1 switches abruptly from central differencing to one-sided differencing when the condition (2.34) is violated. In some schemes, the switch from central to one-sided differencing is made smooth by using a weighted average of a one-sided and a central difference operator. The weight on the central difference is close to one when o(t,x)? > |u(t,x)|A2, but decreases smoothly to zero as o(t, 2)? /|u(t, x)| tends to zero. While it is unclear whether a smooth transition to upwinding is truly important (the convergence order is typically not improved over straight upwinding), Duffy [2000] suggests that the class of exponentially fitted schemes (see Duffy {2000} and Stoyan [1979]) may be quite robust in derivatives pricing applications. In some finance applications, multi-dimensional PDEs might arise where o(t,x) =0 for one of the underlying variables; see for instance Section 2.7.5. While upwinding techniques still apply here, we note that specialized methods exist with better (O(A?)) convergence, should they become necessary. See, for instance, d’Halluin et al. (2005] for details on the so-called semi-Lagrangian methods. 2.7 Option Examples In our discussion so far, we have assumed that options are characterized by a single terminal payoff function g(x) and a set of spatial boundary64 2 Finite Difference Methods conditions determining the option price at the boundaries of the x-domain. In reality, many options are more complicated than this and may involve early exercise decisions, pre-maturity cash flows, path dependency, and more. In this section, we provide some relatively straightforward examples of such complications and how to modify the basic finite difference algorithm to deal with them. More examples will be provided later, in the context of specific fixed income securities. 2.7.1 Continuous Barrier Options We have already touched upon the concept of an up-and-out knock-out option, an option that expires worthless if the x-process ever rises above a critical level H. As we described, we here must simply solve the PDE (2.1) on a domain [M, H], where M represents the lowest attainable value of the process x(t) on [0,7]. The boundary condition at the upper boundary is then dictated to be V(t, H) = 0, ie. of the Dirichlet type. We can generalize this to allow both “up” and “down” type barriers, and to perhaps give a non-zero payout (a “rebate”) at the time the barrier(s) are hit (provided this happens before the option maturity). Specifically, if we have a lower barrier at H, an upper barrier of H, a time-dependent lower rebate function of f(t), and a time-dependent upper rebate function of f(t), we must dimension our spatial grid {2j}5 to have ro =H, tm41 = H, and we then simply impose the Dirichlet boundary conditions V(t, co) = f(t) and V(t,¢m+1) = F(t). See (2.10) and the definition of 9 for the algorithm required to incorporate such Dirichlet boundary conditions into the finite difference scheme. In practice, barrier options sometimes involve time-dependent barriers, possibly with discontinuities. For instance, step-up and step-down barrier options will have piecewise flat barriers that increase (step-up) or decrease (step-down) at discrete points in time. Extension of the finite difference algo- rithm to cover step-up and step-down options is relatively straightforward. As an illustration, consider a zero-rebate up-and-out single-barrier option where the (upper) barrier is flat, except for a discontinuous change at time T* H*. We set the z-domain of our finite difference grid to a € [M,H], with M a probabilistic lower limit, as defined above; accord- ingly, our spatial grid would be {aj }7"G', where x = M and tm4i = H. In preparation for the shift in barrier levels at time T*, we make sure that one level in the spatial grid — say 2,41, k < m, — is set exactly at the level H*. Similarly, we make sure that one level in the time grid is set exactly to T*. Starting at time T, we then iterate backwards in time by repeated solution of m-dimensional tri-diagonal systems of equations, at each step integrating a prescribed rebate function by supplying the Dirichlet boundary condition V(t,’m+1) = 0. The moment we hit T*, the PDE now only applies to the smaller region (M, H*], covered by the reduced spatial grid {«;}4*} with2.7 Option Examples 65 p41 = H*. From T* back to time 0, the backward induction algorithm then involves only k-dimensional tri-diagonal systems of equations, with the Dirichlet boundary condition V(t, r4.41) = 0. Spatial nodes above p41 correspond to zero option value and can be ignored”. Modification of the algorithm outlined above to handle more than two barrier discontinuities is straightforward. We can extend our definition of barrier options even further by making the topology of “alive” and “dead” regions more complicated. At time t, assume for instance that the PDE applies in an “alive” region of x € L(t) and a rebate function R(t, x) that applies in the “dead” region D(t) = B\L(t). Assume that we discretize the problem on a single rectangular fit finite difference grid spanning the spatial domain {M,M], where M and M are set such that the alive regions are covered, up to probabilistic limits (if necessary). Given option values at time t;41, we then only need to run the basic matrix equation (2.18) for values in our grid {«;} that lie inside L(t;). This requires scaling down the dimension of the matrix A as needed, and providing the relevant boundary conditions (given through R(t;,7;)) at the boundary (or boundaries) of L(t;). The parts of the spatial grid that lie outside of L(t,) can be directly filled in with values provided by the rebate function R. Notice that, if possible, the spatial grid should be set such that the boundaries of L(t;) are contained in the mesh; this will likely require us to use the techniques outlined in Section 2.4. If the alive region has the simple form L(t) = [a(t), A(t)] for smooth deterministic functions a and £, an alternative to the scheme above is to introduce a time-dependent transformation that straightens out the barriers, allowing us to return to the standard finite difference setup where the PDE applies to a single rectangular (t,2) domain. One possible transformation involves using a spatial variable of x — a(t) B(t) — a(t)’ which transforms the curved x-barriers a(t) and A(t) into flat y-barriers at y = 0 and y = 1, respectively. The linearity of the transformation (2.41) makes it easy to work with; see Tavella and Randall [2000] for details and a discussion of extensions to multi-dimensional PDEs and to barriers with discontinuities. y=y(t,2) = (2.41) 12 An obvious twist to the algorithm involves using different spatial grids over [0,7] and [2", 7], allowing for more flexibility in node placement. In this case, values computed by backward induction must, at time 7, be interpolated from one z-grid to another. The interpolation rule should be at least third-order accurate; see the discussion in Section 2.7.3.66 2 Finite Difference Methods 2.7.2 Discrete Barrier Options The barrier options considered in Section 2.7.1 are continuously monitored, in the sense that the barrier condition is observed for all times in a given interval. In practice, monitoring the barrier condition continuously can be impractical, and it may instead only be imposed on a discrete set of dates T, < Tz <...< Tx, with Te 0. For the sake of concreteness, let us consider a discretely monitored up-and-out option with a constant barrier H. For a continuously monitored up-and-out barrier option it would suffice to solve the PDE on a domain x € [M, H], where M is a probabilistic lower limit. This is, however, no longer the case for a discretely monitored option where we need to allow the value function to “diffuse” above the barrier levels between dates in the monitoring set {Ti }{<,. To allow for this, we discretize the PDE on a larger domain x € [M,M], M > H. We can determine M probabilistically by determining a confidence interval for how far above the barrier x(t) can rise between monitoring dates. For instance, for the Black-Scholes PDE (2.3), assume that max;=9,...,.« (Tk —Tk-1) = Ar. Conditioned on x(t) = H, the probability that x(t + Ar) exceeds La =H+ (« - 5”) Ar +a0V/ Ar is (—a). As in Section 2.1, we recommend setting 17 = xq for values of a somewhere between 3 and 5. To properly capture diffusion between barrier observation dates, we should also dimension the time grid of the finite difference scheme such that multiple time steps (at least two or three, say) are taken between observation dates. All observation dates {Tk}{Q, should obviously be contained in the time grid. Between barrier observation dates, we solve our PDE by the standard finite difference algorithm outlined in Section 2.2.4, as always imposing cither an asymptotic Dirichlet condition at « = MW or a condition on the a-derivatives of the value function. At each barrier observation time T,,, we must impose a barrier jump condition V(Te-,2) =V(Tet,2) ecu, REL Ke, (2.42) where the notation T,+ was introduced in Section 1.10.1 to denote the limit T, +e for ¢ | 0. This merely states that all values V(Th,a:) are zero for a > H, consistent with the definition of an up-and-out option. In our finite difference scheme, we incorporate this jump condition by simply interpreting the vector V(T,) as found by regular backward induction as V(7,+) and then replacing a ~ A T V(T.+) = (Vile)... Fn(Te)) with2.7 Option Examples 67 ~ ~~ ~ T V(Tk-) = (Tet oycrrys+s Pm(Te+)Ux,, <1) before continuing the algorithm backwards from Tp. The jump condition (2.42) will generally produce a discontinuity in V as a function of x, around the barrier level H. If we use Crank-Nicolson time-stepping, it will then be prudent to employ a fully implicit scheme for the first few backwards time steps (Rannacher stepping) past each barrier observation date T,. As discussed in Section 2.5, ideally this should be combined with a smoothing algorithm acting on V(Z_—) or, perhaps more conveniently, a shift of the spatial grid such that A lies exactly mid-way between two spatial nodes in the grid. We round off by noting that the discussion above for an up-and-out option easily extends to more complicated discrete barrier options, including those with time-varying barrier levels and rebates. For instance, assume that an option involves upper and lower time-varying barriers of H(t) and H(t), respectively, as well as a time- and state-dependent rebate of R(t, x). In this case, we simply replace the jump condition (2.42) with V(i-.2) = V (Tet. 2) ry cxeH(t)} + R(T, ©) (eesti + leecner.))) , and otherwise proceed as above. We note that time-dependent barriers will typically require flexibility in setting the spatial grid, as there are now multiple critical x-levels to consider. The discretization in Section 2.4 can obviously assist with this. 2.7.3 Coupon-Paying Securities and Dividends Many fixed-income securities are coupon-bearing and involve periodic trans- fer of a cash amount between the buyer and the seller. This can easily be incorporated into a finite difference grid, through a jump condition. Specifi- cally, consider a security that pays its owner a single cash amount of p(T*, x) at time T* R. We dimension our time grid such that T* is contained in the grid, and then apply at time T* the condition V(T*~,2) = V(T*+,2) + r(T*, (2.43) This simply expresses that V will decrease by an amount p immediately after p is paid (and thereby no longer contained in V). In a finite difference algorithm, (2.43) is incorporated by replacing V(T*+), as found by regular backward induction, with oa a a T vrr-) = (Aas) 4 p(T*, 1) y-.-; Vm (T*+) +(T",2m))68 2 Finite Difference Methods before continuing the algorithm backwards from T*. Extensions to multiple coupons are trivial. In some cases a derivative security does not itself pay coupons, but is written on a security that does. This involves no particular complications, except for the case where payments may affect the state variable underlying the PDE. For instance, consider the classical case of a stock paying a dividend: at the time of the dividend payment, the stock jumps down by an amount equal to the dividend payment. For a model that uses the stock price (or a transformation of the stock price) as the state variable x, a dividend payment at time T* would thus be associated with a discontinuity in the state variable, 2(T*+) = x(T*—) ~ d(T*,x(T*—)), where d is the magnitude of the jump!’. As long as the dividend-payment does not come as a surprise (ie., at a random time), it must already be contained into the option price at T*~, and will have no price effect as we move forward from T*~ to T*+. We can express this continuity restriction through yet another jump condition V(T*-,2) =V(T"+,2-d(T*,2)). (2.44) See Wilmott et al. [1993] for more discussion. Implementation of (2.44) in a finite difference grid proceeds as follows. First, we use regular backward induction to establish V(T*+) = (ar ys Pm(4))" = (V(r +21), PT*+,2%m)) . Then we write V(T*-) = (¥ (T*+,2, —d(T*,21)),...,0 (T* 4.2m — a(T*,2%m))) The values V;(T*+, R, finite difference grids are ideal for pricing of such securities. Let us first consider a Bermudan option with exercise opportunities restricted to the finite set {Z,}{,. The Bellman principle (1.67) in Section 1.10 can, as shown there, be expressed as a simple jump condition V(Tp—,) = max (V (Th+,2),h(Tk,2)), k=1,...,K, (2.45) which can be incorporated into a finite difference solver precisely the same way as in previous sections. The condition (2.45) will result in a kink in the value function around the level of x at which we shift from the hold region into the exercise region. If Crank-Nicolson time-stepping is used, one should ideally apply smoothing on the finite difference value vector V(T*—), particularly around the kink. If exercise can take place continuously (that is, American-style) on a given time interval, a crude way to incorporate this into a finite difference grid is by simply applying (2.45) to every point in the time grid of the finite difference scheme. By not specifically imposing the partial differential inequalities (see Section 1.10.1), this algorithm, however, will generally only be accurate to first order in the time step, even if a Crank-Nicolson scheme is used; see Carverhill and Clewlow {1990} for a proof. As American-style exercise is rarely used in fixed income markets, we shall not pursue this issue further but just point out that a number of schemes exist to restore second- order time convergence to finite difference pricing of American options, see, e.g., Forsyth and Vetzal [2002]. 2.7.5 Path-Dependent Options Finite difference methods are normally limited to Markovian problems where dynamics are characterized by SDEs and where payouts are simple deterministic functions of the underlying state variables. A number of options, however, have terminal time T payouts that depend not only on the state of x at time T, but on the entire path {x(t), t € (0, 7]}. In general, such options must be priced by Monte Carlo methods (see Chapter 3), but exceptions exist. Indeed, barrier and American options can be considered path-dependent options, yet, as we have seen, can still be priced in a finite difference grid. Even stronger path-dependence can sometimes be handled, through the introduction of new state variables to the PDE. To give an example, consider a path-dependent contract whe: terminal payout at time T can be written as > the 441f h represents the value of a derivative security that has no closed-form pricing formula, it may be necessary to estimate this function by backward induction in the finite difference grid itself. Such a “preprocessing” step is typically straightforward to execute.70 2 Finite Difference Methods V(T) = 9 (a(T),1(7)), (2.46) where J is a path integral of the type t I(t) = [ h(a(s)) ds, (2.47) 0 for some deterministic function h. For instance, if h(x) = x, we say that the option is a continuously sampled Asian option. For the payout (2.46) we have V(t) = V(t, x(t), I(t)) where x(t) satisfies the SDE (2.2) and dI(t) = h(a(t)) dt, I(0) =0. From the backward Kolmogorov equation, it follows that V(t, z, I) solves x + ult, at + So(t,2) Se V + h(a) ie =r(t2)V, (248) subject to the terminal condition V(T,2z,I) = aan There are several complications with this PDE. First, it involves two spatial variables, x and I, requiring the use of a two-dimensional PDE solver. Second, the PDE contains no second-order derivative in the variable J, i.e. it is convection dominated in the I-direction. We have discussed methods to handle the latter issue in Section 2.6.1 and will turn to address the former in Section 2.9. Another complication is the fact that the term h(x) multiplying 0V/OI may be of a different order of magnitude than the other coefficients in (2.48), increasing the difficulty of solving the equation numerically. We refer to Zvan et al. [1998] for a more detailed discussion of PDEs of the type (2.48). In practice, it is rare that a continuous-time integral such as (2.47) is used in an option payout. Instead, one normally samples the function h(x(t)) only on a discrete set of dates, i.e. we replace I(T) with n UL) = So h(a(T)) (Ts -T-1), i=l where Typ < T; <... < Tj, is a discrete schedule, with Ty = 0 and T, = T. Informally, we now have aI(t) = 6(T; - t)-h(a(Ti)) (Ti-Ti-1), 1(0) = 0, (2.49) where 6(-) is the Dirac delta function. In a PDE setting, we incorporate a process such as (2.49) through appropriate jump conditions, writing V(Tj-,2, 1) = V(Tit, 2,1 + h(x) (T; -Ty-1))- (2.50) In the same fashion as for discrete dividends (Section 2.7.3), the jump condition enforces continuity of the option price across the dates where [2.7 Option Examples 7 gets updated. The condition is applied at each date in the discrete schedule, i =1,...,n; in between schedule dates (where now dI(t) = 0), we solve the etal av av eV = t. = V, ae tolisalSe + Solty LE = r(t,aNv, which has no term involving J. When the J-direction is discretized in, say, my different values, the solution scheme thus involves solving m, different one-dimensional PDEs backward in time; the solutions of these my PDEs exchange information with each other at each date in the schedule, in accordance with (2.50). As was the case for cash dividends, implementation of (2.50) will normally require support from an interpolation scheme, to align the (x-dependent) jumps in J with the knots of the discretized I-grid used in the finite difference scheme. See, e.g., Zvan et al. [1999] or Wilmott et al. [1993] for further details. An application of this idea in the context of interest rate derivatives is given in Section 18.4.5. On rare occasions — basically when the homogeneity condition V(nz,XI,t) = X7V(z,1,t), A,n > 0, holds — it is possible to make a change of variables or a change of probability measure that will reduce (2.48) or its discrete-time version to a one-dimensional PDE; see e.g. Rogers and Shi [1995] or Andreasen [1998] for the case of various Asian options. Section 18.4.5 demonstrates one such method, sometimes called the method of similarity reduction, for pricing of “weakly path-dependent” securities, including certain callable interest rate derivatives where the notional accretes at a stochastic coupon rate (see Section 5.14.5 for definitions). 2.7.6 Multiple Exercise Rights Certain financial products with early exercise rights allow the holder to exercise more than once. Such “multi-exercise” options are relatively rare, but the so-called chooser cap (also known as a flezi-cap) is occasionally traded and constitutes a good example for describing how to handle multi- exercise options in a PDE setting. Let there be given a set of L possible exercise dates, T; < Tz <... < T, and assume that we have the right to exercise no more than | times, with | < L. Provided that we exercise at time T,, in a chooser cap we are paid’® ($(T;) — K)+, where S(-) is some interest rate index and K is the strike. Clearly, we would never exercise at time T; unless S(7;) > K, but how much larger than K the rate S(T;) needs to be to trigger optimal exercise is not obvious, and must at least depend on i) how many of our / exercise opportunities we have already used up at time T;; and ii) how much value is lost by using (rather than postponing) one of the remaining exercise opportunities. 18We have ignored a day count scaling constant in the payout. Also, in most cases payment takes place at time 141, rather than at ‘/j; such a payment delay can be handled by a discount operation.72 2 Finite Difference Methods While the question of how to exercise optimally on a chooser cap may appear quite complex, it is surprisingly easy to implement in a finite difference setting by combining techniques from Sections 2.7.4 and 2.7.5 above. The key to the method is to introduce an additional state variable J to keep track of how many exercise opportunities are left. Assume that all interest rates are functions of a Markov state variable x(-), and let therefore V(t, x, I) denote the value of the chooser cap at time ¢, given x(t) = x and given that there are still J exercise opportunities left. Notice that the variable J can only take 1 +1 distinct values: 0,1,...,0; notice also that V(t,x,0) =0 for all t and 2, since J = 0 corresponds to the situation where there are no exercise opportunities left. Additionally, at the terminal time T;, we clearly have V(Tr, 2,1) =(S(TL,e) — K)*, 1 =1,2,.0.51, (2.51) where we have written S(T;) = S(T,,x) to emphasize the deterministic dependence of S on the state variabl For given dynamics of x(t), starting with the terminal conditions in (2.51), we may roll the / different value functions V(-,x, J), J = 1,2,...,l, back through time in standard finite difference manner. At each time T;, i=1,...,L—1, jump conditions similar to (2.45) must be applied, for all 1=1,2..., V(Ti-,@, I) = max (V(Tj+, 2, 1), V(Ti+, x, I ~ 1) + (S(Tj, x) — K)*). Notice that these conditions simply express that exercise is optimal only if the exercise value (the cap payout plus the value of a chooser cap with one less exercise opportunity) exceeds the hold value (the non-exercised chooser cap). Once we have rolled all the way back to t = 0, the chooser cap value at time t = 0 may be identified as V(0) = V(0, x(0), 1). We should note that the “chooser” or “flexi” feature can be added to securities other than caps (and floors). For instance, in Section 19.5 we study the so-called flexi-swap, another security with multiple embedded exercise rights. 2.8 Special Issues In this section, we briefly show a few techniques that may come in handy for certain applications. 2.8.1 Mesh Refinements for Multiple Events As discussed in Section 2.1, the domain of the state variable x is often determined as an exact or approximate confidence interval for the random variable x(T’), where T is the final time of interest for a particular valuation problem we want to solve. Given the number of desired spatial steps in the2.8 Special Issues 73 scheme, the discretization step in z-direction is then obtained by dividing the size of the confidence interval by the number of steps. Similarly, the discretization step in t-direction is typically obtained by dividing T by the number of desired time steps. This is a standard procedure for building a simple rectangular mesh, and it works well if the derivative we wish to value does not have any “interesting” features between the valuation time 0 and the final time T (e.g., for a simple European option). However, as should be evident from the examples in Section 2.7, many real-life derivative securities are characterized by a multitude of events during their lifetimes, all of which must be adequately captured in the PDE scheme. It is not hard to see that a grid dimensioning scheme based solely on the last event date may yield inappropriate mesh resolution at earlier dates. To make the discussion above concrete, let us consider the example of a Bermudan option (see Section 2.7.4) with two exercise dates, T; and T2. Assume that 0 < Ty < To, ie. that the first exercise date is much closer to the valuation date than the second (and last) one. Also assume that there is a decent chance that the option actually will be exercised at time Tj, making it important to capture to good precision the value of the option expiring at T;. Now, if we build our mesh based only on the distribution of the state variable z(T>) at time To, there would typically be too few t-points in the interval [0,7]. Also, the x-direction discretization step would be too large compared to the range of possible values of the state variable x(T,) at time T), i.e. the z-grid would be too coarse for the process x(-) on the time interval (0,7;]. Both issues would typically lead to a large discretization error in the finite difference stepping of the option over the time period {0,T;], leading to problems with accuracy in values and risk sensitivities. The issue of the sparsity of the time grid is fairly easy to deal with, as we are free to add extra points to the time grid before time T;. This by itself, however, will not solve precision problems, as the space step remains large. Any proper solution should, of course, come in the form of refining both the t- and a-grids at the same time. One possible way of refining the «-discretization is to abandon the usage of a single rectangular (t,x)-domain, and instead link together different equidistant rectangular meshes for different periods in the life of the deriva- tive. These mesh “blocks” would generally increase in spatial width with time and would connect to each other via an interpolation scheme. To be more specific, let us assume, as in Section 2.1, that the state variable x(-) is the logarithm of the stock in the Black-Scholes model and is given by (2.4), with the PDE to solve given by (2.3). We extend our simple two-period example above to a derivative with K times of interest, 0 < T; <... < Tx; these times could be specified as an additional input into valuation, or derived from the trade description (e.g. they could represent the exercise dates for a Bermudan option, or the knock-out dates for the discretely-monitored barrier option of Section 2.7.2). Suppose we are given values of m and n, and now wish to construct the mesh for the time period [T},_1, Th], by using74 2 Finite Difference Methods the same time and space steps A¥, AK as would be used in the standard scheme of Section 2.1 for a derivative security with the terminal payoff at Tk. That is, having fixed the cutoff a we would set Ab =T/n, Ak = 2a0V/Tk/(m +1). (2.52) Then the rectangular, equidistant mesh for the time period [T,~1,T7p] is given by Cee aR, Ta HAAR, oh = hag HAE, (2.53) where |-| denotes the integer part of a real number and (see (2.4)) hin = (0) + (" - 3”) Th — 00 Tp (2.54) Note that in reality we would want to make sure that the point TJ), is also in the mesh for the time period [T},1,7;], even though for simplicity of notations we did not reflect it in (2.53). It is also useful to note that the total number of time points is not going to be n, but is actually equal to K So [= Tea/Te) n}, k=l which scales linearly with n. Clearly, if exactly n points were required, a simple adjustment to the definition of the time step in (2.52) could be applied. With a mesh as defined above, when arriving at time Tj, in a backward induction scheme the solution V(Tj,+) would be discretized on the z-grid att me a . To solve the PDE backwards over the time period [Ty~1,Tx], we would need to resample it on the different x-grid {e*}7™5'. As with interpolation across dividends (Section 2.7.3), simple cubic ” interpolation would be a good choice here. Specifically, one would fit a cubic spline to the values V(Tk,2}""), 7 = 0,...,m+ 1, and then calculate V(Tp,2*), j =0,...,m-+1, by valuing the spline at the required grid points. The “interpolated mesh” scheme above is rather intuitive and straight- forward, but it does suffer from the need to do interpolation work that could slow down the PDE (especially in dimensions higher than 1 and/or for a large number of interface points K). Also, it is not entirely clear how inter- polation will affect stability and convergence properties of the PDE. Finally, linking the interface mesh geometry to the trade specifics (such as exercise dates) may not be ideal from the point of view of designing an efficient valuation flow in a risk management system. These considerations lead us to an alternative approach that relies on non-equidistant discretization as developed in Section 2.4. The idea of this method is to use non-uniform2.8 Special Issues 75 discretization to concentrate more points, both in time and space, around the initial point t = 0, = 2(0). Clearly many ways of achieving this are possible — below we present a simple scheme we have used with good results. We define K, the user input, to be the number of spatial refinement. levels (with K = 2 or 3 typically used), and r, another user input, to be a time scaling constant (typically r = 4). If T is the final horizon for valuation, we then introduce times 0=%) k=1 k=L where anink ak + jak, 2h = okt 4 7k, ‘6 And, as advised earlier, adding trade event dates that fall into this period — although we do not reflect this in our notations for simplicity.76 2 Finite Difference Methods and m= |= | Re |= = that | _ ar At This distribution of space points results in an z-grid that is more dense around the point 2 = (0) than at the edges. It is worth noting that with only one refinement level K = 1, the standard rectangular uniform mesh sized by the terminal distribution of the state variable is recovered. 2.8.2 Analytics at the Last Time Step In cases where the dynamics of underlying PDE variables are tractable, one naturally wonders whether finite difference methods could somehow be improved by incorporating analytical results into the scheme. Here, and in the next section, we discuss two simple ideas. Suppose that we are faced with the problem of pricing a contingent claim with terminal boundary condition g(a(T)), where 2(¢) is a Markovian process with known Arrow-Debreu state prices: G(t,2;s,y) = E2 (6 (2(s) — y) en FE nwa) de} p(y = 2) , s>t Assume also that the claim in question involves a jump condition at time 0 < T* < T (but no jump conditions between J* and T). If our finite difference grid is {xj}, we can now use a series of m + 2 outright convolutions to compute V(I*,2;) = [erate g(y)dy, 7 =0,1,...,.m+1. (2.55) If we are lucky (i.e., if both g and G are sufficiently simple), then the integral on the right-hand side may be known in closed form for all values of x;. If not, we can always perform a series of numerical integrations, the total cost of which is typically!” O(m?), i.e. more expensive than the typical O(m) cost of a single time step in a finite difference method. There are several reasons why we may want to perform the numerical integrations nevertheless. First, the convolution expression (2.55) is exact, as it is based on the true transition density. Second, if the gap between T* and T is large, an ordinary finite difference grid would need to roll back from T to T* using multiple time steps n*, at a total cost of O(n*m); if n* is of the same magnitude as m, the computational effort of the convolution scheme would be comparable to that of a finite difference grid. Third, for discontinuous payouts, the integration in (2.55) will have a naturally smoothing effect, similar to (but often better than) the continuity correction method of Section 2.5.2. The smoothing 'TThere are exceptions. For instance, if fast Fourier transform (FFT) methods are applicable, the cost may be reduced to O(m In(m)). See Section 8.4 for details.2.8 Special Issues 17 effect is discussed in more detail in Section 23.2.4 and is also demonstrated below, in Figure 2.2, where we have continued our investigation of the 3 year digital option considered earlier in Section 2.5.3. Since the model used in Figure 2.2 is ordinary Black-Scholes and g(a) = 1¢2>1}, the integrals in (2.55) can here be computed in closed form from (2.33). Fig. 2.2. 3 Year Digital Option Price —— Analytical Smoothing 0.30 3 Straight Crank-Nicolson 0.25 + —_—_——. 5 15 25 35 45 Grid Points in Asset Direction (m ) Notes: Finite difference estimates for the Black-Scholes price of a 3 year digital option. All contract and model parameters are as in Figure 2.1. Time stepping is performed with an equidistant grid containing n = 50 points. Spatial discretization in log-space is equidistant, as described in the main text; the number of grid points (m) is as listed on the a-axis of the figure. The “Straight Crank-Nicolson” graph shows the convergence profile for a pure Crank-Nicolson finite difference grid. The “Analytical Smoothing” graph shows the convergence profile for a Crank-Nicolson finite difference grid starting at 7 = 2.5 years, with the terminal boundary condition set equal to a 0.5 year digital option price (as in (2.55)). The theoretical value of the option is 0.4312451. In principle, we could continue rolling back from T* (through, possi- bly, jump conditions at earlier times) by performing convolutions, rather than solving finite difference grids. In practice, this rarely leads to improve- ments over a finite difference grid, unless the densities and payoffs are quite simple!§. Moreover, in many cases we may not have exact Arrow-Debreu 18For simple densities (especially Gaussian), special-purpose methods exist to compute convolutions rapidly, typically involving payoff approximations through piecewise polynomials or other simple functions. We do not cover these methods in73 2 Finite Difference Methods prices, only approximate ones based on, say, a small-time expansion (see, e.g., Section [Link]). In this case, a one-time convolution may be safe especially if 7 —T* is small — whereas repeated convolutions may lead to unacceptable biases. 2.8.3 Analytics at the First Time Step The idea in Section 2.8.2 of replacing the finite difference stepping with analytical integration is even easier to apply over the first, rather than the last, time step. Suppose 7 is the first “interesting” time for a given derivative security, ie. there might be a jump condition at time T* but none over the time interval {0,7*]. Then, rather than stepping the finite difference scheme from T* to 0, we can perform a single integration to calculate the value V(0,x(0)) of the derivative at time zero from the discretized values {V(T",x5)}""5' of the derivative at time T* (using the same notations as in Section 2.8.2), 0);T*,y) V(T*,y) dy, V(0,x(0)) = [cu R where V(T*,y) is interpolated (using cubic splines, say) from the values {V(T*,a5)}"4" on the grid. If the integral is computed numerically — as is most often the case — the numerical cost is often comparable with that of the finite difference stepping because only one value V(0,2(0)) is required at time 0, not the whole slice. While there are typically no numerical cost savings that arise from using integration over the first time step, there are accuracy and stability considerations that favor this approach. We have already seen in Section 2.8.1 that the standard discretization of a PDE often leads to insufficient fidelity in resolving any features of the payoff that are close to today, and numerical integration can be of considerable help in this regard. Moreover, as we discuss in much detail later in Chapter 23, an integration scheme typically allows us to treat discontinuities in the value V(T*, a) arising from the jump condition at time 7* explicitly. If the discontinuity is introduced at the value of the state variable x*, then the integration scheme can (and should) explicitly take this information into account. For example we would write voo,x(0) = [7 G (0,2(0);T*, y) V> (Ty) dy + [60.2079 Hv) dy this book except for a brief mention in Section 11.A. For a representative example see Hu et al. [2006].2.9 Multi-Dimensional PDEs: Problem Formulation 79 and calculate V~ (T*,y) by interpolating the grid values in the time inter- val (—co,*), and V+(T*,y) by interpolating the grid values in (z*,0o), separately. The usefulness of the method is only limited by the availability of the closed-form expression for the time 0 Arrow-Debreu prices G(0,x(0);T”*,-). For some models this is not an issue; for most others, sufficiently close approximations could be obtained in a small-time limit (see e.g. Section [Link] for a typical approach) that can be useful for times T* that are not too large. By a change of measure, we see that. V(0,2(0)) =E (e ha “(T+ 2(T"))) = P(,T*)E™ (V(T*,(T*))), where ET” is the expected value operator under the T*-forward measure Q™”; so we really only need the expression for the density (rather than Arrow- Debreu security prices) of c(T*) under Q?”, either exact or approximate. Finally, we note that while the integration over the first time step can be seen to offer similar advantages to those of the methods in Section 2.8.1, the two approaches are not substitutes for each other, but are complementary. We typically recommend using direct integration over the time step [0, 7}, where T* is the smaller of the time of the first jump condition or the limit of applicability of the approximation to the density of «(T*), and then (if needed) use the methods in Section 2.8.1 over the time interval [T*,T), with T being the final maturity of the option in question. 2.9 Multi-Dimensional PDEs: Problem Formulation We now turn our attention to the numerical solution of multi-dimensional terminal value problems. Let the spatial variable x be p-dimensional, x = (t1,-.-,2p)", and consider the PDE P wv i PV Lunt 5 TFL lb Aga — Mhz) =0, (2.56) h=1l=1 av ot where spp(t,) > 0 and spu(t,x) = si,n(t,2) for h,l = 1,...,p. The PDE is assumed subject to the terminal value condition V(T,r) = g(x), g: R? > R. From the results in Chapter 1, we recognize that the PDE provides the solution to the expectation '°9One of the functions V~(7",y), V* (1, y) is often known analytically and for all values of y (rather than sampled on the grid); this is for instance the case for the Bermudan options of Section 2.7.4. ‘Che integration algorithm should obviously take advantage of this.80 2 Finite Difference Methods V(t) = By (ero dg (2(7)) [x(¢) = 2), where the components of x(t) satisfy risk-neutral SDEs of the type dap,(t) = up (t, x(t) dt + on (t,a(t)) dW(t), h=1,...,p. (2.57) Here W(t) is a d-dimensional Brownian motion, pn : [0,T] x R? > R, h=1,...,p, are (scalar) drifts, and op : (0,T] x R? 3 RI*4,h=1,...,p, are d-dimensional (row vector) diffusion coefficients. The PDE coefficients sp, in (2.56) represent the instantaneous covariance matrix for the components of 2(-), Le., Spu(t,2) = on(t,x)or(t, 2)". We assume enough regularity on Hn; Th, 7, and g to ensure that (2.56) has a unique solution. For the purpose of solving (2.56) numerically, we assume that the PDE is to be solved on a (finite) spatial domain in x, « € [M,,My]x...x(M,, Mp), where Mn,Mn, h=1,...,p, are constants either dictated by the contract at hand (barrier options) or found by a suitable probabilistic truncation (see Section 2.1). 2.10 Two-Dimensional PDE with No Mixed Derivatives To illustrate the construction of finite difference discretization of (2.56), we start out with the simple case where p = d = 2 and there are no mixed partial derivatives in the PDE: s),9(t,z) = so,1(t,c) = 0 for all t and a. Probabilistically, the absence of mixed derivatives corresponds to the case where the stochastic process increments dz ;(t) and dx2(t) are independent. Defining 7,(t,2)? = sa,n(t, x), h = 1,2, the PDE to be solved now becomes ov at (Li + L2)V =0, (2.58) where Oo 1 Lr = walt, 2a "as Flt x) aa au 5r(te), h=1,2. Notice that we have divided the term r(t, x) into equal pieces in £; and Lo. To discretize (2.58) in x, introduce grids x, € {xi'}M$' and x2 € {a }n25'. To simplify notation, assume these grids are equidistant such that 2}! = M, + j1Ay and 2? = My + jog. Let Vj,,5,(t) £ V(t, 27, 222). We define discrete central difference operators as before Vij ttsja (t) = Vir 1,50 (t) [Link] salt) = Be f t Sx2Vjnjn(t) = tee Ben2.10 ‘Two-Dimensional PDE with No Mixed Derivatives 81 and [Link](t) — [Link](t) + Vin —ayjo(t bese, Vjngp(t) = (int 152 (t) ‘ug ja ranlt) i joa t) ~ Vjcsjo(t) + Viacio—1(6) BnaeaVj() = dee — Tag 4 Mae, These operators, in turn, give rise to the discrete operators 7 4 1 2 1 Ln = Unt, 2)52, + 5In(t@)"bx,2, — 572), h=1,2, where « is constrained to take values in the spatial grid. A Taylor expansion shows that this operator is second-order accurate (compare to Lemma 2.2.1), (Li + Lo)V(t,2) = (£1 + £2) V(t.) + O (A? + 43). 2.10.1 Theta Method Turning to a theta-style time discretization, consider first proceeding exactly as in Section 2.2.3. Assuming equidistant time spacing A;, we get for the period [t:, tis] ( 0; (A + 2,)) Vj sia(ta) = (141 =0)4: (Er + £2) Visa (ten) + ef", where et =O (4: (43 + AB + Log} Ae + 4?) , and where it is understood that Ly and Lo are to be evaluated at (t,x) = (t)7 (0), ct 23?) with t!71 (0) defined as in (2.14). If Vj, 5, (t) = V(t,24', 3?) is a finite difference approximation to V},,;.(t), we thus get the scheme (1-04 (Er + £2) Piaslte) = (14 0-4 (G+ 22) Fh saltina) (2.59) to be solved for the mymy interior points Vj, j)(ti), jr = 1,--.,ma, de = 1,...,me, given the values of Vine (t.4.1), and given appropriate boundary conditions at 7; = 0, j1 =m, +1, jo =0, and jg =m24+1. The scheme (2.59) represents a system of linear equations in m,mz un- knowns {¥},,;,(ti)}. When written out as a matrix equation (which requires us to arrange the various Voasjo (ta) in some order in a (7m ,mg)-dimensional vector), the matrix to be inverted is sparse but, unfortunately, no longer tri-diagonal. Solution of the system of equations by standard methods (e.g., Gauss-Jordan elimination or LU decomposition) is out of the question due82 2 Finite Difference Methods to the size of the matrix?°. We can proceed in two ways: either we use a specialized sparse-matrix solver; or we attempt to redo the discretization (2.59) to make it computationally efficient. We personally prefer the second approach and shall outline one method in the next section. As for the first approach, we simply note that a good iterative sparse solver should be able to solve (2.59) in order O((mmz)°/4) operations. See Saad [2003] for concrete algorithms. 2.10.2 The Alternating Direction Implicit (ADI) Method The ADI method is an example of a so-called operator splitting method, where the simultaneous application of two operators (here £; and Lo) is split into two sequential operator applications. To illustrate the idea, set 9 = 3 (Crank-Nicolson scheme) in (2.59) and approximate (1 - sau (E4 &)) ns ( 7 jac) (: 7 pada) , (2.60) (1454 (é:+2)) x (1+ 54:21) (1+ 542). (2.61) It is easy to see*! (and to verify, by a Taylor expansion) that the operators on the right-hand sides of these approximations have the same order truncation error as do the left-hand sides, namely O(A;(A? + A3 + A?)). To the order of our original scheme, no accuracy is gained or lost in using the right-hand sides of (2.60)-(2.61). What is gained, however, is a considerable improvement in computational efficiency, originating in the fact that the resulting scheme 1 > 1 a\s (: - pati) (: = jada) Vj, 50 (ta) 1 = 1 a\s = (14 54d) (14+ $42) Fatt) (269) can be split into the system lye L,s\s5 ( - jae) Ving = ( + jae) Vin cin (tetas (2.63) l,a\s lis ( - a) Vir jo (ti) = ( + ye) Upasins (2.64) ?°Recall that the solution of a general linear system with mim2 unknowns is an O(m?m3) operation. For, say, m1 and mz in the order of 100, this would involve around 1,000,000 times more work than what is required for a one-dimensional (tri-diagonal) scheme (O(m)). °1To those versed in operator notation, we notice that the right- and left-hand sides both approximate, to identical order, exp(+0.54:(£1 + L2)).2.10 Two-Dimensional PDE with No Mixed Derivatives 83 where we have introduced an intermediate value Uj,,5.. The advantage of this decomposition is the fact that in each of (2.63) and (2.64), there is only one operator on the left-hand side, leading to simple tri-diagonal equation systems. To formalize this, first define UP = (Uigias Ur,iay-- Umacia) + Then, for a fixed value of jg we can write for the first step t, i: Y (raat (ise ‘)) ue? = Me (2344) , (2.65) where A? is an (m; x m;)-dimensional tri- “diagonal matrix of the same form as (2, 11) (to get Aj’, basically freeze r2 = x3? and substitute 4, and ‘1 for 4 and o in the definition of the one- dimensional matrix A). The my-dimensional vector M3? has components Me, vip J = 1,..-,™1, given by tir tti l,o\s5 ME, (A344) = (2 + 54:L2) ee V5, jo —1(tiga) + 5. jaVind-sa(tint) = Five 1 + (1- Jone) Masaltonds (2.68) where we have defined 7 batt > tiga tt sha Ba (nm (44 at 0f) + Ante (HES ot of A tigr tt Pa turtti jf 5 sno # Sf (m (Hatt a att) + 545r tt tad ok 7 For known values of 7(t:41), (2.65) defines a simple tri-diagonal equation system which can be solved for Uj? in O(m) operations. Repeating the procedure above for jg = 1,...,mg allows us to find U;,;, for all jy = 1,..-,™1, jo =1,..., me, at a total computational cost of O(m mz). Turning to the second step of (2.63)-(2.64), we first fix j; and define VP) = (Pra, FralOs-+sFramal) In the same fashion as earlier, we can then write (1-54 (#3 *)) VR (t;) = Me (= *) : (2.67) where A} is an (mg X mg)-dimensional tri-diagonal matrix and where the right-hand side vector now has components84 2 Finite Difference Methods ty. t. 1 ra . Mi, (4) 7 ( + pe) Uji J2= For brevity we omit writing out the Mi, (which will be similar to (2.66)), but just notice that the right-hand side of (2.67) is known after the first step of the ADI algorithm (above) is complete. For a given value of #1, we can solve the tri-diagonal system (2.67) for ve (t;) in O(mg) operations. Looping over all m, different values of j;, the full matrix of time t; values Vinsie (ta), ji=1l,...,11, jz =1,..-, ma, can then be found at a total computational cost of O(inymg). The scheme outlined above is known as the Peaceman-Rachford scheme. As is the case for all ADI schemes, the scheme works by alternating the directions that are treated fully implicitly in the finite difference grid: in the first step, the 21-direction is fully implicit and the «2-direction is fully explicit, and in the second step the order is reversed. In effect, both spatial variables end up being discretized “semi-implicitly”, i.e. similar to a Crank-Nicolson scheme, resulting in convergence order is O(A? + A3 + A?). We emphasize, however, that whereas a direct application of the Crank-Nicolson scheme will involve (if an efficient sparse-matrix solver is used) a computational cost of O((mimz)*/4) per time step, the computational cost of the Peaceman- Rachford ADI scheme is only O(m1mg). A (tedious) von Neumann analysis reveals that the scheme is A-stable, but, like the Crank-Nicolson scheme, not strongly A-stable. While the Peaceman-Rachford scheme is a classical example of an ADI scheme, there are many others. For instance, consider a theta-version of the Douglas-Rachford scheme: +) Ma. Vusaltest), — (2.68) (i - 0A:L2) Vyasa (ta) = Usa gs — OAL Vj, ia (tina), (2.69) (Q 7 0A:L:) Uys = (a + (1-9) AL + AL, where we understand that in Lt and Lo the PDE coefficients are to be evaluated at time ti*1(6). Again, notice how the scheme consists of two steps, each involving the solution of tri-diagonal sets of equations along only one of the 2}- or x2-directions. The computational cost thus remains at O(m,mg). It can be shown that the convergence order of this scheme is O(A} + A3 + Lyoz4y4e + A?) and it is A-stable for @ > 5, and strongly A-stable for @ > 4. By elimination of U;,,;, we note that the unsplit version of the Douglas-Rachford scheme is (a - 04,£1) (a - 0ArL2) Vpn ja (ta) = ((1- 04:21) ( 0A £2) + AL + A.L2) Va rin(tist)- It is not difficult to see that this approximates (2.59) to second order.2.11 ‘Two-Dimensional PDE with Mixed Derivatives 85 2.10.3 Boundary Conditions and Other Issues The fact that ADI schemes reduce to solving sequences of matrix systems identical to the ones arising in the one-dimensional case is convenient, in the sense that many of the issues we have encountered for one-dimensional finite difference grids (oscillations, stability, convection dominance, etc.) and their remedies (smoothing, non-equidistant discretization, upwinding, etc.) carry over to the ADI setting with only minor modifications. Con- sider for instance the issue of applying spatial boundary conditions along the edges of the (:c},22) domain, which we have so far not discussed. As for the one-dimensional PDEs, the most convenient way to express such boundary conditions is typically by imposing conditions on derivatives, like OV (t, 29, 2?)/Ox? = V(t, 29, e3?)/Ax, and so forth. For the Peaceman- Rachford scheme, say, such conditions can be incorporated directly into (2.65) and (2.67) by altering the matrices Aj? and Aj}, as well as the boundary elements of Mj’ and M3, in the manner outlined in Section 2.2.1. If instead we wish to impose Dirichlet boundary conditions, we need to add corrective terms to the tri-diagonal systems, as in (2.19). To complete the first part of the split scheme, this then requires us to establish what boundary terms are needed for the intermediate quantity Uj,,;., ie. we must define U,,9 and U5, ,mz+1 for jy = 1,..., m1, as well as Uo,;, and Um, 41,3 for jg = 1,...,me. While U;,,;, is a purely mathematic construct, sometimes it is adequate to think of Uj,,,, as a proxy for V;,;, evaluated at ¢}+!(@), which obviously makes determination of boundary conditions straightforward. For maximum. precision, however, we should use the ADI equations themselves to express the boundary conditions of U directly in terms of boundary conditions for V(t,) and V(t,41). Here, the Douglas-Rachford scheme is particularly easy to deal with, as a rearrangement of (2.69) directly relates U,, j, to Vas ga (ts) and Vj,.y,(te41)s Unga = (1 AL 2) Vy, salts) + OAL utes) The Peaceman-Rachford scheme requires some further manipulations to express U in terms of V(t;) and V(t,41); see Mitchell and Griffiths [1980] for the details. 2.11 Two-Dimensional PDE with Mixed Derivatives Consider now the case where the 2-dimensional PDE (2.58) has a mixed partial derivative, Vv My (Cr+ Le Lia) V =, (2.70) where £; and £o are as in (2.58), and where86 2 Finite Difference Methods 2 2 Lir= sialtit)a a olt,z)n (t.e)ralt") aS aay (2.71) The quantity p(t, ) is the instantaneous correlation between the processes x(t) and x9(t) in (2.57), ie. p(t,) € [-1,1]- The presence of £;,2 prevents a direct application of the ADI methods in Section 2.10.2, since the mixed operator £;,2 is not amenable to operator splitting. We shall demonstrate two ways to overcome this problem: a) orthogonalization of the PDE; and b) predictor-corrector schemes. 2.11.1 Orthogonalization of the PDE The idea here is to introduce new variables y;(t,21,22) and yo(t, 21, 22) such that the PDE loses its mixed derivative term when stated in terms of these variables. To demonstrate this idea, assume first that p(t,x), y(t, x), and y(t, x) are all functions of time only and independent of x. Then define, say, w(t, 1,02) = 21, (2.72) y2(t) mn(t) where we must assume that y(t) 4 0 for all t. yo(t, 21, £2) = —p(t) 21 +22 Sa(t)t1 +22, (2.73) Lemma 2.11.1. Consider the PDE (2.70) subject to the terminal value condition V(T, 2) = g(x). Define y = (yi,y2)" and u(t,y) = V(t,2). With the variable change defined in (2.72)-(2.73), v satisfies wvioiy Ov oy, Ov 1g Oy at Fao aT tHE + ond) ag 1 2 207u +3 (1 ~ p(t)?) y(t) ag r(t,yn,y2—a(t)y1) = 0, (2.74) Y2 where Hi (t,y) © mn (t, 21,22) = ma (tyr, y2 — alt), (2.75) da(t Blt) & On, + alten (t,22,20) + Ha (ts 21,22) da(t = By, + a(Qut(ty) +ue(tave—alOnn). (276) The equation (2.74) is subject to the terminal value condition v(T, 41, y2) = 9(21, 02) = 9(y1,y2 — a(T)y)- Proof. While the result can be established by the usual mechanics of ordinary calculus, we will take the opportunity to show how stochastic calculus can2.11 ‘I'wo-Dimensional PDE with Mixed Derivatives 87 also conveniently prove results of this type. Going back to the processes underlying the PDE (see (2.57)), we write dxy(t) = p(t, x) dt + y1(t) dWi(t), (2.77) dea(t) = pir(t,x) dt + y2(t) (0 (t) dWy(t) + VT — ple)? dWa(t )) (2.78) for independent scalar Brownian motions W(t) and W(t); this is easily seen to generate the correct correlation p(t) between x, and x. An application of Ito’s lemma then shows that the processes for y; and yo are dyi(t) = dara (ty = palt, 2) dt + ne) dWi(t), dy(t) = 20 ay(t) dt + a(t) (t,2) dt + a(t)yn(t) dW (t) + nalts) dt + alt) (old) aWi(t) + VT pO? aWa(t)) 7 (= ay(t) + a(t)un(t,2) + molt, )) at + a(t) V1 — p(t)? dWo(t) With the definitions (2.75)-(2.76), this becomes simply dys(t) = wy (t y(t) dt + n(t) dWi(t), (2.79) dya(t) = u3 (t,y(t)) dt + y(t) /1 — p(t)? dWalt). (2.80) Equations (2.79)-(2.80) define a Markov SDE in y;(t) and yo(t) where, importantly, the Brownian motions on y(t) and yo(t) are now independent. Writing V(t,x) = v(t,y), it then follows immediately from the backward Kolmogorov equation (see Section 1.8) that v satisfies the PDE (2.74). 0 Through the chosen transformation (2.72)-(2.73), our original PDE has now been put into a form where we can immediately apply the ADI schemes outlined in Section 2.10.2. In performing the orthogonalization of the PDE in Lemma 2.11.1 we relied on p(t,x), y1(t,), and 72(t,r) all being independent of x. This can often be relaxed. Consider for instance the case where p(t,x) = p(t), yi(t,c) = y(t, 21), and y(t, x) = yo(t, x2); here the correlation p is still assumed deterministic, but we now allow for some (though not full) x- dependence in 7 and 72. Assuming that 7; (t,71) > 0 and y(t,x2) > 0 we can introduce new variables 2 (t.21) = | Ea (2.81) 29 (ty) = / sana" (2.82) Applying Ito’s lemma to (2.77)-(2.78) we see that88 2 Finite Difference Methods _ On(tti) 1 a(t, x) data = (~ [adn t een 1 dy (t,21) 2 an dt+dW,(t) (2.83) and _ Oy2(t, 22) 1 H(t, x) dealtm) = (- ot Aol, ea)" a 7y2(t, 2) ~; me) dt -+ p(t) dWy(t) + JT =p dWe(t). (2.84) 2 As we assumed that 7 (t,21) > 0 and 72(t, 2) > 0, the functions z; and zq are increasing in «, and 22, respectively, and are thereby invertible. As such, we can rewrite (2.83)-(2.84) in the more appealing form dzy(t, 1) = yj (t, 21, 22) dt + dW, (t), dza(t,c1) = 43 (t, 21, 22) dt + p(t) dWi(t) + V1 — p(t)? dWa(t). Through the transformation (2.81)-(2.82), we have reduced our original system to one where the coefficients on W1(t) and W(t) are no longer state-dependent, similar to the case that lead to Lemma 2.11.1. We can now proceed with another variable transformation, as in (2.72)-(2.73), to orthogonalize the system and prepare it for an application of the ADI method. While the orthogonalization method outlined here can be very effective on a range of practical problems, it suffers from a few drawbacks. Most obviously, the method is not completely general and requires a certain structure on the parameters of the PDE. Another drawback is that the introduction of a time- dependent transformation on one or more variables (Lemma 2.11.1) often makes the alignment of the finite difference grid along (time-independent) critical level points in x-space impossible. Also, the introduction of terms like y;da(t)/dt in the drift of yo (see (2.76)) can be problematic, particularly if the functions 7,(¢) and y(t) are not smooth. For instance, it is not unlikely that y,da(t)/dt will locally be of such magnitude that upwinding will be necessary to prevent oscillations; see Section 2.6.1. Further, we note that inversion of the transformations (2.81)-(2.82) will not always be possible to perform analytically and may require numerical (root-search) work, complicating the scheme and potentially slowing it down. Finally, as we shall highlight in future chapters, maintaining the “continuity” of a numerical scheme with respect to input parameters is of critical importance for the smoothness of risk sensitivities. Such continuity is difficult to ensure if complicated transformations are applied to model variables. So, in the end, we recommend formulating the PDEs in terms of financially meaningful variables, avoiding excessive transformations, and relying on methods such as developed in the next section when dealing with mixed derivatives and other numerical complications.2.11 ‘I'wo-Dimensional PDE with Mixed Derivatives 89 2.11.2 Predictor-Corrector Scheme In this section we shall consider a completely general method for handling mixed derivatives in two-dimensional PDEs. While a bit slower than the method outlined in Section 2.11.1, it does not involve any variable transfor- mations and, by extension, does not suffer from the drawbacks associated with such transformations. As a first step, consider the discretization of the mixed derivative 0?V/dx,0x2. There are a few possibilities (see Mitchell and Griffiths [1980]), but we shall just use Se122 Vir sie (t) = Sirs 5ca Vina (€) _ Vatiieti(t) — Visine) = Vinita) + Vinnie lt) ia (2.85) Extensions to non-equidistant grids follow directly from (2.27) and the relation x, 02Vj; jo (t) = bx, 5x2 Vj1,j.(t). As we have not encountered mixed difference operators before, for completeness we show the following lemma. Lemma 2.11.2. For the discrete operator (2.85) we have V(t, 24', a) 2 2 Bm, Be2 +0 (Aj + A3). Sxyz2V ju ja(t) = Proof. A Taylor expansion of V(t, ) around the point « = (x}',2?)" gives snag a ta i clasps ala Ata. Vj —1jeti(t) = Vj ie (t) — ae “a a 1 oe 1, 4 BV (1, #8V BA ang + 2 aaa * A little thought then shows that Vin etottlt) — Virtig—1) — Vir~tieta(t) + Vj 1-1) = 4A, Ao +0 (A}Ap + A143), Vv 0x1 0x990 2 Finite Difference Methods as error terms of order Af, A$, and A?A3 will cancel. The result follows. Oo Equipped with (2.85), we can approximate the operator £;,9 in (2.71) as LiaVi elt) £0 (tata) n (tat 2) ae (tat 2) dna Vi salt) which is accurate to order O(A? + A3). The first easy way to modify our ADI scheme to incorporate Lis is to treat the mixed derivative fully explicitly. In the Douglas-Rachford scheme (2.68)-(2.69), for instance, we thus modify the right-hand side of the first step as follows: (1 = 04:£1) [Link] = (14-0) AL, + ArLa + Ara) Be (tins)s (2.86) (1 = @A:L2) Belts) = Vine — 0AeE2V, jatar: (2.87) The addition of Lio this way clearly preserves the ADI structure of the scheme which will continue to involve only sequences of tri-diagonal linear equations. However, having, in effect, only a one-sided time-differencing of the mixed derivative teri will lower the convergence order of the time step to O(A,), irrespective of the choice of 6. To change the time at which the mixed operator Ly, is evaluated, consider using a predictor-corrector scheme, where the results of (2.86)— (2.87) are re-used in a one-time”? iteration. Specifically, we write, for some A € [0,1], Predictor: (1-44:L21) Uf, = (14 (1 OAL + Le + ALi 2) Vin (tins), (2.88) (16422) UP), = 00, - PAL, joltiar). (2-89) Corrector: (a - 9ArL:) 2) = (1 + (1- OAL, + ALo + (1-NAL a), Tr iv(tvt) +ALL 2,,, (2.90) (1 = 0AcL2) Vy gots) = 24, — OALE2V, 9, (to41)- (2.91) *2We can run the iteration more than once if desired, but a single iteration will normally suffice.2.12 PDEs of Arbitrary Order 91 Notice how the Douglas-Rachford scheme is first run once, in (2.88)— (2.89), to yield a first guess (a “predictor” ), ue), for the time t; value Vj, .j2(t,). In a second run of the Douglas-Rachford scheme, in (2.90)-(2.91), this guess is used as a “corrector” to affect the time at wet Li, gis evaluated, by applying this operator to (1 — A), jo(tis1) + AU?,; when 4 = 4 we effectively center the time-differencing of the mixed term. The scheme now relies on three intermediate variables, U\),,, US)... and 2... The combined predictor-corrector scheme above (in a slightly less general form, with A; = Ag) was suggested by Craig and Sneyd [1988]. It can be shown that the scheme has convergence order 0 ((Ar + 40)? + 1yozg} Ae + Leg} de + 42), so second order convergence in the time domain is still achievable by setting 6 == }. The scheme will be A-stable for @ > 3 and 4 » ins) V(tisa), h=2 h=1l=h4+1 (1 7 6A,£2) U® =U —GA,LP (tr), 1-0A,L,) U®) =U?) — GAL V (tis). P Pp Corrector: ( -9A,£;) 20 = 7 > =A; (@" +(1-AL14+ 0 Lr h=2 P P P P +1-)Y ins) (tina) FADS SO La, h=ll=htl h=1l=h41 (2 - 0A.£2) Z% = ZY ~ 9A,LoV (tis), (1-0A:L,) P(t) = 2° — AZ, P (ti41). With mp, points in the xp-direction, h = 1,...,p, the computational cost of the predictor-corrector scheme is O([],_, hn). For p < 3, sufficient conditions for A-stability are 9 > 3 and 3 < A < 8. For p > 4, sufficient conditions are 9 < 3 and p-1 d ov, Yn-# 4 ohva NO): where N(0,1) is a standard Gaussian distribution and-4 denotes convergence in distribution'. Purther, if we define "Recall that a sequence of variables X, with cumulative distribution func- tions F, converge in distribution to a random vi ble X with distribution F if limta-soo Fa(w) = F(e) for all x € R at which F(2) is continuous.3.1 Fundamentals 97 then also _ Yn-bu Sn/Vn Define the Gaussian percentile u, as ®(u,) = 1 — y, where @ is the Gaussian cumulative distribution function. From Theorem 3.1.2, and from the definition of convergence in distribution (see footnote 1), the probability that the confidence interval [V(t) - wyy2° 8n/ V2, V(t) + uy/2 > Sn/Vr] (3.4) fails to include the true value V(t) approaches y for large n. Here 4N(0,1). with the quantity s,/./n known as the standard error. For given y, the rate at which the confidence interval for V(t) contracts is O(n~#). This is relatively slow: to reduce the width of the interval by a factor of 2, n must increase by a factor of 4. On the other hand, we notice that the (asymptotic) convergence rate only depends on n, not on the specifics of the g;’s. In particular, if g(T) = g(X(T)) where X is p-dimensional, the asymptotic convergence rate is independent of p. As we shall see shortly, in most applications the work required to generate samples of g(X(T)) is (at most) linear in p. 3.1.1 Generation of Random Samples At the most basic level, the Monte Carlo method requires the ability to draw independent realizations of a scalar random variable Z with a specified cumulative distribution function F(z) = P(Z < z), where P is a probability measure. On a computer, the starting point for this exercise is normally a pseudo-random number generator, a software program that will generate a sequence of numbers uniformly distributed on (0, 1] (i.e. from 2/(0, 1). Press et al. [1992] list a number of generators producing sequences of uniform numbers uj, u2,.-. from iterative relationships of the form Tizi = (al; +c) mod(m), itn = Tiga /m. The externally specified starting point Ip is the seed of the random number generator. In this so-called general linear congruential generator, the choice of the multiplier a, the modulus m, and the increment ¢ must be done