Introduction to Stochastic Calculus

Transcription

1 Introduction to Stochastic Calculus Director Chennai Mathematical Institute Introduction to Stochastic Calculus - 1

2 The notion of Conditional Expectation of a random variable given a σ-field confuses many people who wish to learn stochastic calculus. What does given a σ-field mean? Thus we begin with a discussion on Conditional Expectation. Introduction to Stochastic Calculus - 2

3 Let X be a random variable. Suppose we are required to make a guess for the value of X, we would like to be as close to X as possible. Suppose the penalty function is square of the error. Thus we wish to minimize E[(X a) 2 ] (1) where a is the guess. The value of a that minimizes (1) is the mean µ = E[X ]. Introduction to Stochastic Calculus - 3

4 Let Y be another random variable which we can observe and we are allowed to use the observation Y while guessing X, i.e. our guess could be a function of Y - g(y ). We should then choose the function g such that E[(X g(y )) 2 ] (2) takes the minimum possible value. When Y takes finitely many values, the function g can be thought of as a look-up table. Assuming E[X 2 ] <, it can be shown that there exists a function g (Borel measurable function from R to R) such that E[(X g(y )) 2 ] E[(X f (Y )) 2 ]. (3) Introduction to Stochastic Calculus - 4

5 H = L 2 (Ω,F,P) - the space of all square integrable random variables on (Ω,F,P) with norm Z = E[Z 2 ] is an Hilbert space and K = {f (Y ) : f measurable, E[(f (Y )) 2 ] < } is a closed subspace of H and hence given X H, there is a unique element in g K that satisfies Further, E[(X g(y )) 2 ] E[(X f (Y )) 2 ]. E[(X g(y ))f (Y )] = f K. Introduction to Stochastic Calculus - 5

6 For X H, we define g(y ) to be the conditional expectation of X given Y, written as E[X Y ] = g(y ). One can show that for X,Z H and a,b R, one has E[aX + b Z Y ] = ae[x Y ] + b E[Z Y ] (4) and X Z implies E[X Y ] E[Z Y ]. (5) Here and in this document, statements on random variables about equality or inequality are in the sense of almost sure, i.e. they are asserted to be true outside a null set. Introduction to Stochastic Calculus - 6

7 If (X,Y ) are independent, then Y does not contain any information on X and E[X Y ] = E[X ]. This can be checked using the definition. If (X,Y ) has bivariate Normal distrobution with means zero, variances 1 and correlation coefficient r, then from properties of Normal distribution, it follows that X ry is independent of Y. It follows that g(y) = ry is the choice for conditional expectation in this case: E[X Y ] = ry. Introduction to Stochastic Calculus - 7

8 Now if instead of one random variable Y, we were to observe Y 1,...,Y m, we can similarly define E[X Y 1,...Y m ] = g(y 1,...Y m ) where g satisfies E[(X g(y 1,...Y m )) 2 ] E[(X f (Y 1,...Y m )) 2 ] = measurable functions f on R m. Once again we would need to show that the conditional expectation is linear and monotone. Introduction to Stochastic Calculus - 8

9 Also if we were to observe an infinite sequence, we have to proceed similarly, with f, g being Borel functions on R, while if were to observe an R d valued continuous stochastic process {Y t : t < }, we can proceed as before with f, g being Borel functions on C([, ),R d ). In each case we will have to write down properties and proofs thereof and keep doing the same as the class of observable random variables changes. A different approach helps us avoid this. Introduction to Stochastic Calculus - 9

10 For a random variable Y, (defined on a probability space (Ω,F,P)), the smallest σ-field σ(y ) with respect to which Y is measurable (also called the σ-field generated by Y ) is given by σ(y ) = { A F : A = {Y B}, B B(R) }. An important fact: A random variable Z can be written as Z = g(y ) for a measurable function g if and only if Z is measurable with respect to σ(y ). In view of these observations, one can define conditional expectation given a sigma field as follows. Introduction to Stochastic Calculus - 1

11 Let X be a random variable on (Ω,F,P) with E[ X 2 ] < and let G be a sub-σ field of F. Then the conditional expectation of X given G is defined to be the G measurable random variable Z such that E[(X Z) 2 ] E[(X U) 2 ], U G -measurable r.v. (6) The random variable Z is characterized via E[X 1 A ] = E[Z1 A ] A G (7) It should be remembered that one mostly uses it when the sigma field G is generated by a bunch of observable random variables and then Z is a function of these observables. Introduction to Stochastic Calculus - 11

12 We can see that information content in observing Y or observing Y 3 is the same and so intutively E[X Y ] = E[X Y 3 ] This can be seen to be true on observing that σ(y ) = σ(y 3 ) as Y is measurable w.r.t. σ(y 3 ) and Y 3 is measurable w.r.t σ(y ). Introduction to Stochastic Calculus - 12

13 We extend the definition of E[X G ] to integrable X as follows. For X, let X n = X n. Then X n is square integrable and hence Z n = E[X n G ] is defined and Z n Z n+1 as X n X n+1. So we define Z = limz n and since E[X n 1 A ] = E[Z n 1 A ] A G monotone convergence theorem implies E[X 1 A ] = E[Z1 A ] A G and it follows that E[Z] = E[X ] < and we define E[X G ] = Z. Introduction to Stochastic Calculus - 13

14 Now given X such that E[ X ] < we define E[X G ] = E[X + G ] E[X + G ] where X + = max(x,) and X = max( X,). It follows that E [ ] [ ] E[X G ]1 A = E X 1A. A G (8) The property (8) characterizes the conditional expectation. Introduction to Stochastic Calculus - 14

15 Let X,X n,z be integrable random variables on (Ω,F,P) for n 1 and G be a sub-σ field of F and a,b R. Then we have (i) E[aX + b Z G ] = ae[x G ] + b E[Z G ]. (ii) If Y is G measurable and bounded then E(XY G ) = Y E(X G ). (iii) X Z E(X G ) E(Z G ). (iv) E [ E[X G ] ] = E[X ]. Introduction to Stochastic Calculus - 15

16 (v) For a sub-sigma field H with G H F one has E [ E[X H ] G ] = E[X G ] (vi) E[X G ] E[ X G ] (vii) If[ E[ X n X ] then ] E E[X n G ] E[X G ]. Introduction to Stochastic Calculus - 16

17 Consider a situation from finance. Let S t be a continuous process denoting the market price of shares of a company UVW. Let A t denote the value of the assets of the company, B t denote the value of contracts that the company has bid, C t denote the value of contracts that the company is about to sign. The process S is observed by the public but the processes A,B,C are not observed by the public at large. Hence, while making a decision on investing in shares of the company UVW, an investor can only use information {S u : u t}. Introduction to Stochastic Calculus - 17

18 Indeed, in trying to find an optimal investment policy π = (π t ) (optimum under some criterion), the class of all investment strategy must be taken as all processes π such that for each t, π t is a (measurable) function of {S u : u t} i.e. for each t, π t is measurable w.r.t. the σ-field G t = σ(s u : u t). In particular, the strategy cannot be a function of the unobserved processes A,B,C. Introduction to Stochastic Calculus - 18

19 Thus it is useful to define for t, G t to be the σ- field generated by all the random variables observable upto time t and then require any action to be taken at time t (an estimate of some quantity or investment decision) should be measurable with respect to G t. A filtration is any family {F t } of sub-σ- fields indexted by t [, ) such that F s F t for s t. A process X = (X t ) is said to be adapted to a filtration {F t } if for each t, X t is F t measurable. Introduction to Stochastic Calculus - 19

20 While it is not required in the definition, in most situations, the filtration (F t ) under consideration would be chosen to be {F Z t : t } where F Z t = σ(z u : u t) for some process Z, which itself could be vector valued. Sometimes, a filtration is treated as a mere technicality. We would like to stress that it is not so. It is a technical concept, but can be a very important ingredient of the analysis. Introduction to Stochastic Calculus - 2

21 For example in the finance example, we must require any investment policy to be adapted to F t = σ(s u : u t). While for technical reasons we may also consider H t = σ(s u,a u,b u,c u : u t) an (H t ) adapted process cannot be taken as an investment strategy. Thus (F t ) is lot more than mere technicality. Introduction to Stochastic Calculus - 21

22 A sequence of random variables {M n : n 1} is said to be a martingale if E[M n+1 M,M 1,...,M n ] = M n, n 1. An important property of martingale: E[M n ] = E[M 1 ]. The process D n = M n M n 1 is said to be a martingale difference sequence. An important property of martingale difference sequence: E[(D 1 + D D n ) 2 ] = E[(D 1 ) 2 + (D 2 ) (D n ) 2 ] as cross product terms vanish: for i < j, E[D i D j ] = E [ D i E[D j M 1,M 2,...,M j 1 ] ] = as {D j } is a martingale difference sequence. Introduction to Stochastic Calculus - 22

23 Thus martingale difference sequence shares a very important property with mean zero independent random variables: Expected value of square of the sum equals Expected value of sum of squares. For iid random variables, this property is what makes the law of large numbers work. Thus martingales share many properties with sum of iid mean zero random variable, Law of large numbers, central limit theorem etc. Indeed, Martingale is a single most powerful tool in modern probability theory. Introduction to Stochastic Calculus - 23

24 Let (M n ) be a martingale and D n = M n M n 1 be the martingale difference sequence. For k 1, let g k be a bounded function on R k and let G k = g k (M,M 1,...,M k 1 )D k. Then it can be seen that G k is also a martingale difference sequence and hence N n = n k=1 g k(m,m 1,...,M k 1 )D k is a martingale. Note that the multiplier is a function of M,M 1,...,M k 1 and NOT of M,M 1,...,M k. In the later case, the result will not be true in general. Hence E[(G 1 + G G n ) 2 ] = E[(G 1 ) 2 + (G 2 ) (G n ) 2 ] Introduction to Stochastic Calculus - 24

25 The processes {N n } obtained via N n = n k=1 g k(m,m 1,...,M k 1 )(M k M k 1 ) is known as the martingale transform of M. We have seen that the martingale transform {N n } is itself a martingale and so E[N n ] =. Introduction to Stochastic Calculus - 25

26 Note that E(N n ) 2 = = n k=1 n k=1 E[(g k (M,M 1,...,M k 1 )(M k M k 1 )) 2 ] E[(g k (M,...,M k 1 ) 2 E[(M k M k 1 ) 2 σ(m 1,...,M k 1 )] and hence if g k is bounded by 1, then E(N n ) 2 = n k=1 E[(M k M k 1 )) 2 ] n E[E[(M k M k 1 ) 2 σ(m 1,...,M k 1 )] k=1 Introduction to Stochastic Calculus - 26

27 A digression: Consider a discrete model for stock prices: Say S k denotes the price on k th day. Let us assume for the sake of clarity, that the price changes at 2pm and trading is allowed at 11am. Thus, on k th day, the number of shares that an investor could buy can use the information S,S 1,S 2,...,S k 1 (and other publicly available information). On k th day the trisection takes place at the price S k 1. Let us denote by ξ k the number of shares that an investor decides to buy. Then ξ k should be a function of S,S 1,S 2,...,S k 1, say ξ k = g k (S,S 1,S 2,...,S k 1 ). Introduction to Stochastic Calculus - 27

28 The collection of functions {g k : 1 k N} can be considered as an investment strategy. For such a strategy, on k th day the gain (loss) would be ξ k (S k S k 1 ). Thus the net gain (loss) from the strategy {g k : 1 k N} would be N G N = ξ k (S k S k 1 ). k=1 For simplicity let us assume that rate of interest is, else we can replace S k by S k = S k (1 + r) k, where r is the rate of interest. Introduction to Stochastic Calculus - 28

29 Suppose there exist functions {g k : 1 k N such that P(G N ) = 1 and P(G N > ) >. Such a strategy is called an arbitrage opportunity. In a market in equilibrium, arbitrage opportunities cannot exist for if they did, all investors would like to follow the same, multiple times over, disturbing the equilibrium. Introduction to Stochastic Calculus - 29

30 If {S k } were a martingale, then the gains process {G n } is a martingale transform and hence a martingale, and so E[G N ] =. As a result, P(G N ) = 1 implies P(G N = ) = 1 and so arbitrage opportunity cannot exist. Suppose there exists a probability Q such that P(E) = if and only if Q(E) =. Such a Q is called an Equivalent probability measure. Introduction to Stochastic Calculus - 3

31 For an Equivalent probability measure Q P(E) > if and only if Q(E) > and P(E) = 1 if and only if Q(E) = 1. So a strategy {g k } is an arbitrage opportunity under P if and only if it is an arbitrage opportunity under Q. Thus even if the stock price is not a martingale under the model probability measure P, but there exists an equivalent probability measure Q under which it is a martingale (Q is then called an Equivalent Martingale Measure - EMM), then arbitrage opportunities do not exist. Introduction to Stochastic Calculus - 31

32 Interesting thing is that when each {S k : k N} takes finitely many values, the converse is also true- No Arbitrage holds if and only if EMM exists. The result is also essentially true in great generality in continuous time too, one needs some technical conditions. This result is called Fundamental theorem of Asset Pricing. This shows a deep relation between Martingales and Mathematical Finance. Introduction to Stochastic Calculus - 32

33 A process (M t ) adapted to a filtration (F t ), ( for s < t, F s is sub-σ field of F t and M t is F t measurable) is said to be a martingale if for s < t, E [ M t F s ] = Ms. For = t < t 1 <... < t m, in analogy with the discrete case one has E [ (M tm M t ) 2] = E [ m 1 j= (M t j+1 M tj ) 2] Under minimal conditions one can ensure that the paths t M t (ω) are right continuous functions with left limts (r.c.l.l.) for all ω Ω. So we will only consider r.c.l.l. martingales. Introduction to Stochastic Calculus - 33

34 For any integrable random variable Z on (Ω,F,P), the process Z t = E[Z F t ] is a martingale. This follows from the smoothing property of conditional expectation: for G H Here H = F t and G = F s. E [ E[Z H ] G ] = E[X G ]. Introduction to Stochastic Calculus - 34

35 One of the most important result on Martingales is: Doob s Maximal inequality For an r.c.l.l. Martingale (M t ), for all λ >, and P ( sup t T M t > λ ) 1 λ 2 E ( M 2 T E ( sup t T M t 2 ) 4E ( M 2 T ). ) Introduction to Stochastic Calculus - 35

36 It can be shown that for any t n 1 k= (M t(k+1) M tk ) 2 A t n n where A t is an increasing r.c.l.l. adapted process, called the quadratic variation of M. Further, when E[MT 2 ] < for all T, then ) n 1 k= ((M E t(k+1) M tk ) 2 F tk B t n n n where B t is also an increasing r.c.l.l. adapted processand M 2 t A t and M 2 t B t are martingales. Introduction to Stochastic Calculus - 36

37 The quadratic variation process (A t ) of M is written as [M,M] t and B t is written as M,M t. M,M t is called the the predictable quadratic variation of M. When M has continuous paths (t M t (ω) is continuous for all ω) then M,M t =[M,M] t. For Brownian motion (β t ), [β,β] t = σ 2 t and β,β t = σ 2 t, while for a Poisson process (M t ) with rate λ, N t = M t λt is a martingale with then [N,N] t = λt and N,N t = M t. Introduction to Stochastic Calculus - 37

38 A stopping time τ is a [, ] valued random variable such that {ω : τ(ω) t} F t, t} The idea behind the concept: Whether the event {τ t} has happened or not should be known at time t, when one has observed {X u : u t.}. If (X t ) is an continuous adapted process then τ = inf{t : X t θ} is a stopping time : for any θ R. For example, {τ 7} has happened iff {sup s 7 X s θ}. Introduction to Stochastic Calculus - 38

39 In real life we are used to talking in terms of time in reference to time of occurence of other events, such as meet me within five minutes of your arrival or let us meet as soon as the class is over, or sell shares of company X as soon as its price falls below 1. Thse are stopping times Of course a statement like buy shares of company X today if the price is going to increase tomorrow by 3 is meaningless. One can talk but cannot implement. The time specified in the statement is not a stopping time. The time at which an americal option can be exercised has to be a stopping time as the decision to exercise the option at a given time or not has to depend upon the observations upto that time. Introduction to Stochastic Calculus - 39

40 Martingales and stopping times: Let (M t ) be a martingale, τ be a stopping time and (X t ),(Y t ) be r.c.l.l. adapted processes and Y be bounded. 1 (X t ) is a martingale if and only if E[X τ ] = E[X ] for all bounded stopping times τ 2 N t = M t τ is a martingale. 3 U t = Y t τ ( Mt M t τ ) is also a martingale. A martingale (M t ) is said to be a square integrable martingale if E[M 2 t ] < for all t <. Introduction to Stochastic Calculus - 4

41 A process (X t ) is said to be locally bounded if there exists an increasing sequence of stopping times {τ n } increasing to such that X n t = X t τ n is bounded for every n. It can be seen that every (X t ) such that X is bounded and (X t ) is continuous is locally bounded. Likewise, an adapted process (M t ) is said to be a local martingale (or locally square integrable martingale) if there exists an increasing sequence of stopping times {τ n } increasing to such that M n t = M t τ n is a martingale (is a square integrable martingale) for every n. The sequence {τ n } is called a localizing sequence. Introduction to Stochastic Calculus - 41

42 Let M be a local martingale with localizing sequence {τ n }. Thus E[M t τ n F s ] = M s τ n. So if M is bounded then M is a martingale. Other sufficient conditions for a local martingale to be a martingale are [ ] E sup s T M s < [ E M T 2] <, T < [ ] E M T 1+δ <, T <, for some δ >. Introduction to Stochastic Calculus - 42

43 Let M be a [, ) valued local martingale with localizing sequence {τ n }. Then since E[M t τ n] = E[M ] using Fatou s lemma it follows that for all t >, E[M t ] E[M ]. If E[M t ] = E[M ] it is a martingale. Introduction to Stochastic Calculus - 43

44 When (M t ) is a locally square integrable martingale, it can be shown that for any t n 1 k= (M t(k+1) M tk ) 2 A t n n where again A t is an increasing adapted process such that M 2 t A t is a local martingale. A is again called quadratic variation process, written as [M,M] t. One can also show that there is an increasing adapted r.c.l.l. process B such that M n t = M t τ n, M n,m n t = B t τ n. and then B t is written as M,M t. Introduction to Stochastic Calculus - 44

45 Let (M t ) be a locally square integrable martingale with M = and let {τ n } be a localizing sequence such that M n t = M t τ n is a martingale for every n. Then, using E[Mt τ 2 n] = E[ M,M t τ n] = E[[M,M] t τn] and Doob s inequality, one has E ( sup M s 2) 4E[ M,M t τ n] (9) s t τ n as well as E ( sup M s 2) 4E[[M,M] t τ n] (1) s t τ n Introduction to Stochastic Calculus - 45

46 A Brownian motion (β t ) is a continuous process with β = and such that for t 1 < t 2 <... < t m ; m 1; (β t1,β t2,...,β tm ) has a multivariate Normal distribution with mean vector and variance co-variance matrix Σ = ((σ ij )) given by σ ij = min(t i,t j ) It can be checked that (with t = ) β t1 β t,β t2 β t1,...,β tm β tm 1 are independent Normal random variables, with mean and Var(β tj β tj 1 ) = t j t j 1. Introduction to Stochastic Calculus - 46

47 One of the most important properties of a Brownian motion is that (β t ) is a martingale with [β,β] t = β t,β t t = t. Thus, βt 2 t is a martingale and ( ) 2 Qt n = k=1 β k 2 n t β k 1 2 n t t where the convergence is in probability. The Qt n for a fixed n,t is a finite sum, indeed it has j = [2 n t] + 1 terms, where [2 n t] is the integer part of 2 n t since for k such that [2 n t] + 1 < k, k 1 2 n t = t and 2 k n t = t. Introduction to Stochastic Calculus - 47

48 Let Uk n = ( β k 2 n t β ) 2 k 1 2 n t ( k 2 n t k 1 2 n t) Then W n t = [2n t]+1 k=1 U n k = Qn t t and using properties of Normal distribution, it follows that for each n, {U n k : 1 k [2n t]} are i.i.d. random variables, with mean and variance 2 2 2n. Hence and hence Wt n in probability. Variance(Wt n ) ([2 n t] + 1) 2 2t+1 2 2n 2 n converges to zero in probability and so Q n t t With little work, it can be shown that Q n t converges to t, uniformly on t [,T ] for all T < almost surely. Introduction to Stochastic Calculus - 48

49 In fact, for a continuous process (X t ), if X = and if (X t ) and (Y t = X 2 t t) are martingales then (X t ) is Brownian motion (Levy s characterization). For this, continuity of (X t ) is very important. Result not true for r.c.l.l. martinagles. X t = N t t is a counter example where (N t ) is a poisson process with parameter 1. Introduction to Stochastic Calculus - 49

50 Let us fix an observable process (Z t ) (it could be vector valued) with Z t denoting the observation at time t and let F t = σ(z u : u t) denote the filtration generated by observations. A Brownian motion w.r.t. this filtration (F t ) means that β t and β 2 t t are martingales w.r.t. this filtration (F t ). Introduction to Stochastic Calculus - 5

51 Let (β t ) be a Brownian motion w.r.t. a filtration (F t ). There are occasions when one needs to consider limits of expressions that are similar to Riemann-Steltjes sums for the integral t Y s dβ s where (Y s ) is say a continuous (F t ) adapted process: Rt n = k=1 Y s k,n t ( β k 2 n t β k 1 2 n t ) where k 1 2 n s k,n k 2 n. For Riemann-Steltjes integral t Y s dβ s to exist, the above sums (R n t ) should converge (as n ) for all choices of s k,n satisfying k 1 2 n s k,n k 2 n. Introduction to Stochastic Calculus - 51

52 Let Y = β. Let us choose s k,n to be the lower end point in every interval : s k,n = k 1 2 n and for this choice let us denote the Riemann sums Rt n by A n t. Thus A n ( ) t = β k 1 2 k=1 n t β k 2 n t β k 1 2 n t Introduction to Stochastic Calculus - 52

53 Let us now choose s k,n to be the upper end point in every interval : s k,n = 2 k n and for this choice let us denote the Riemann sums Rt n by Bt n. Thus Bt n ( ) = β k 2 n t β k 2 n t β k 1 2 k=1 n t Introduction to Stochastic Calculus - 53

54 Recall while and hence A n t = k=1 β k 1 2 n t B n t = k=1 β k 2 n t B n t A n t = k=1 ( ) β k 2 n t β k 1 2 n t ( ) β k 2 n t β k 1 2 n t ( ) 2 β k 2 n t β k 1 2 n t = Q n t. Since Qt n t, A n t and Bt n cannot converge to the same limit and hence the integral t Y s dβ s cannot be defined as a Riemann-Steltjes integral. Introduction to Stochastic Calculus - 54

55 Indeed, Brownian motion is not an exception but the rule. It can be shown that if a continuous process is used to model stock price movements, then it can not be a process with bounded variation as that would imply existence of arbitrage opportunities. Hence, we cannot use Riemann-Steltjes integrals in this context. The Riemann like sums appear when one is trying to compute the gain/loss from a trading strategy, where the process used to model the stock price appears as an integrator. Introduction to Stochastic Calculus - 55

56 It turns out that in the context of construction of diffusion processes, Riemann sums, where one evaluates the integrand at the lower end point, are appropriate. So taking that approach, Ito defined stochastic integral (in 4s). It so happens that in the context of applications to finance, the same is natural and so Ito Calculus could be used as it is. This was observed in the early 8s. Introduction to Stochastic Calculus - 56

57 Let X t denote the stock price at time t. Consider a simple trading strategy, where buy-sell takes place at fixed times = s < s 1 <... < s m+1. Suppose the number of shares held by the investor during (s j,s j+1 ] is a j. The transaction takes place at time s j and so the investor can use information available to her/him before time t, so a j should be F sj measurable. So the trading strategy is given by: at time t hold f t shares where f s = a 1 {} (s) + m j= a j1 (sj,s j+1 ](s) and then the net gain/loss for the investor due to trading in shares upto time t is given by V t = m j= a j(x sj+1 t X sj t) = t f u dx u Introduction to Stochastic Calculus - 57

58 Now suppose (X t ) is a square integrable martingale with r.c.l.l. paths (right continuous with left limits). Recalling f s = a 1 {} (s) + m j= a j1 (sj,s j+1 ](s), V t = m j= a j(x sj+1 t X sj t). it follows that (V t ) is also a martingale and [V,V ] t = Hence, using (1) [ E sup( t T = t m 1 j= t a 2 j ([X,X ] s j+1 t [X,X ] sj t) f 2 u d[x,x ] u ] [ T f u dx u ) 2 4E fu 2 d[x,x ] u ]. Introduction to Stochastic Calculus - 58

59 The estimate [ E sup( t T t ] [ T ] f u dx u ) 2 4E fu 2 d[x,x ] u for simple integrands f can be used to extend the integral to integrands g that can be approximated by simple functions in the norm X given by RHS above: [ T ] f 2 X = E fu 2 d[x,x ] u Introduction to Stochastic Calculus - 59

60 Let (Y t ) be a bounded adapted continuous process and (X t ) be a square integrable martingale. Let tm n = m n. Then it can be shown that (Y t ) can be approximated in the norm X by simple processes Y n defined by Yt n = Y t n m t1 (t n m, tm+1 n ] (t) m= and hence the stochastic integral t Y u dx u is defined and is given by t Y u dx u = lim n Y t n m t m= ( ) X t n m+1 t X t n m t. The two infinite sums above are actually finite sums for every n,t (if m > tn the terms are zero). Introduction to Stochastic Calculus - 6

61 Also, the limit t Y u dx u = lim n Y t n m t m= ( ) X t n m+1 t X t n m t. is in the sense of convergence in probability, uniformly in t [,T ] (follows from Doob s inequality). Introduction to Stochastic Calculus - 61

62 In other words, for every ε >, as n tends to infinity [ t ) P Y u dx u Y t n m t( ] Xt n m+1 t X t n m t > ε. sup t T m= This shows that the stochastic integral t Y u dx u is a process with r.c.l.l.paths. Further, if X is a continuous process then so is t Y u dx u continuous process. Of course, when X is a process with bounded variation the integral t Y u dx u is defined as the Riemann-Steltjes integral. Introduction to Stochastic Calculus - 62

63 One can directly verify that when X is a square integrable martingale, t Y n u dx u = ) Y t n m t( Xt n m+1 t X t n m t m= is a square integrable martingale. Then convergence of t Y n u dx u in L 2 to t Y u dx u for every t implies that M t = t Y u dx u is also a square integrable martingale. Introduction to Stochastic Calculus - 63

64 One of the important properties of the integral Z t = t Y u dx u so defined is that for a stopping time τ, defining X [τ] t = X t τ, t Y u dx [τ] u = Z t τ. This can be seen to be true for simple integrands and hence for bounded continuous processes Y. Thus if X is a locally square integrable martingale so that for stopping times {τ n } increasing to, X n = X [τn] are square integrable martingales, then Z n = YdX n satisfy Zt n = Zt τ m n, for m n. Introduction to Stochastic Calculus - 64

65 Thus we can define Z t = Zt n for τ n 1 < t τ n and we would have Zt n = Z t τ n. Thus we can define Z = YdX. Thus for a locally square integrable martingale X, for all continuous adapted processes Y we continue to have for every ε >, as n tends to infinity P [ sup t T t Y u dx u m= ) Y t n m t( Xt n m+1 t X t n m t > ε By construction, each Z n is a martingale and hence Z is a local martingale. ]. Introduction to Stochastic Calculus - 65

66 If X were a process with bounded variation paths on [,T ] for every T, of course t YdX is defined as Riemann Steltjes integral and we have P [ sup t T t Y u dx u m= ) Y t n m t( Xt n m+1 t X t n m t > ε ] and then t Y u dx u is a process with bounded variation paths. Introduction to Stochastic Calculus - 66

67 An r.c.l.l. process X is said to be a semimartingale if it admits a decomposition X t = M t + A t, where (M t ) is a locally square integrable martingale and (A t ) is an adapted process whose paths t A t (ω) have bounded variation on [,T ] for every T < for every ω. Such a process A will be called process with bounded variation paths. Introduction to Stochastic Calculus - 67

68 Let X be a semimartingale, with X = M + A, M a locally square integrable martingale and A a process with bounded variation paths. For a bounded continuous adapted process Y, the integral t Y s dx s can be defined as the sum of t Y s dm s and t Y s da s and again one has every ε >, as n tends to infinity P [ sup t T t Y u dx u m= ) Y t n m t( Xt n m+1 t X t n m t > ε Thus the integral YdX does not depend upon the decomposition X = M + A. ]. Introduction to Stochastic Calculus - 68

69 We have seen that for a semimartingale X and a bounded continuous adapted process Y, the integral t Y s dx s can be defined on the lines of Riemann-Steltjes integral, provided we always evaluate the integrand at the lower end point while taking the Riemann sums: t Y u dx u = lim n m= ) Y t n m t( Xt n m+1 t X t n m t. Further, if X is a local martingale, then so is Z t = t Y u dx u and if X is a process with bounded variation paths, then so is Z and as a result, Z is always a semimartingale. Introduction to Stochastic Calculus - 69

70 Question: Just like one extends Riemann integral 1 f (x)dx defined for say continuous functions g on [,1] to Lebesgue integral 1 gdλ for all bounded measurable functions, can we extend t Y s dx s defined for all bounded continuous adapted processes to a larger class of integrands? The natural class of processes is then the class that is closed under bounded pointwise convergence and contains all bounded continuous adapted processes. Introduction to Stochastic Calculus - 7

71 If we think of every process Y t as a function on Ω = [, ) Ω, then the required class is the collection of all P-measurable bounded functions, where P is the σ-field generated by bounded continuous adapted processes. P is called the predictable σ-field and P measurable processes are called predictable processes. Introduction to Stochastic Calculus - 71

72 A left continuous adapted process (Z t ) is predictable, but an r.c.l.l. (right continuous with left limits) adapted process (V t ) need not be predictable. (V t ) is predictable if for all t, V t is F t ε measurable. If (Z t ) is predictable with respect to a filtration (F t ) then for each t, Z t is measurable w.r.t. F t = σ ( s<t F s ). This justifies the use of the name predictable. Introduction to Stochastic Calculus - 72

73 Thus we require a mapping J X from the class of bounded predictable process f into r.c.l.l. processes such that for adapted continuous processes f, J X (f ) t = t fdx (as defined earlier) and if f n converges to f bounded pointwise, then [ ] P sup t T J X (f n ) t J X (f ) t > ε. When X is a semimartingale, then such an extension exists and we denote it as J X (f ) t = t fdx. Introduction to Stochastic Calculus - 73

74 The Bichteller-Dellacherie-Meyer-Mokodobodski Theorem states that for an r.c.l.l. process X, existence of mapping J X from the class of bounded predictable process f into r.c.l.l. processes such that for adapted continuous processes f, J X (f ) t = t fdx (as defined earlier) and if f n converges to f bounded pointwise, then [ ] P sup t T J X (f n ) t J X (f ) t > ε implies that X s a semimartingale. We will not discuss proof of this. Introduction to Stochastic Calculus - 74

75 Let X be a semimartingale. It can be shown that (f,x ) fdx is linear. Important fact: If τ is a stopping time, then g t = 1 [,τ] (t) is predictable (as it is adapted, left continuous). Now for a bounded predictable process (f t ), h t = f t 1 [,τ] (t) is predictable and t h s dx s = t τ f s dx s = f s dx s [τ] where X [τ] s = X s τ. Introduction to Stochastic Calculus - 75

76 Localization: An adapted process (f t ) is said to be locally bounded if there exists an increasing sequence of stopping times {τ n } increasing to such that ft n = f t 1 [,τ n ](t) is bounded for every n. Here, for a stopping times τ, the stochastic intervals [, τ], play the role of compact sets [ K,K] in R. Recall definition of local martingale. Given f, {τ n } as above, for a semimartingale X, let Then for m n Z n t = t f 1 [,τ n ](s)dx s Z n t = Z m t τ n Introduction to Stochastic Calculus - 76

77 Hence we piece together the pieces to define t f s dx s = Z n t for τ n 1 < t τ n. Thus for a semimartingale X and for locally bounded process f, we have defined stochastic integral Z t = t f s dx s. It is an r.c.l.l. process. Moreover, ( Z) t = f t ( X ) t. One can verify this for simple f and then for general f via approximation. Introduction to Stochastic Calculus - 77

78 If X is a square integrable local martingale, then so is Z, if X has bounded variation paths then so does Z and in general, Z is a semimartingale with decomposition Z t = t f s dx s = t t f s dm s + f s da s = N t + B t where X = M + A, M a locally square integrable martingale, A a process with bounded variation paths and N t = t f s dm s is a locally square integrable martingale and B t = t f s da s is a process with bounded variation paths. Introduction to Stochastic Calculus - 78

79 Further if g is a locally bounded predictable process, then t gdz = t gfdx. One can verify this first when f,g are simple functions and then for a fixed g simple one can first verify it for all f and then the general case. Introduction to Stochastic Calculus - 79

80 An r.c.l.l. adapted process Y need not be predictable. However, the process Y defined by Yt = Y t (left limit at t for t > ), Y = is predictable. It is also locally bounded and with tm n = m n one has lim n m= Y t n m t( Xt n m+1 t X t n m t ) t = Yu dx u Introduction to Stochastic Calculus - 8

81 Note that (using b 2 a 2 = 2a(b a) + (b a) 2 ) X 2 t X 2 = = m= m= + ( X 2 t n m+1 t X t 2 ) m t n ) 2X t n m t( Xt n m+1 t X t n m t m= Hence for a semimartingale X, ( Xt n m+1 t X t n m t ( ) t 2 Xt n m+1 t X t n m t X 2 t X 2 2 m= ) 2 X u dx u. Introduction to Stochastic Calculus - 81

82 We denote the limit of 2 m=( Xt n m+1 t X t n m t) by [X,X ]t and note that it is an increasing adapted process. Further we have t Xt 2 = X Xu dx u + [X,X ] t. Introduction to Stochastic Calculus - 82

83 For semimartingales X,Y let us define [X,Y ] by the parallelogram identity [X,Y ] t = 1 4 ([X + Y,X + Y ] t [X Y,X Y ] t ) It follows that [X,Y ] t is a process with bounded variation paths and [X,Y ] t = lim n ) ) m=( Xt n m+1 t X t n m t (Yt n m+1 t Y t n m t Also, it follows that X t Y t = X Y + t Y s dxs + t X s dy s + [X,Y ] t. This is called the integration by parts formula. Introduction to Stochastic Calculus - 83

84 Like Lebesgue integrals, we have a Dominated Convergence Theorem for stochastic integrals: Let (X t ) be a semimartingale and let f n,f be predictable processes such that f n g where g is some locally bounded predictable process and such that f n converges pointwise to f. Then [ t t ] P fu n dx u f u dx u > ε. sup t T Introduction to Stochastic Calculus - 84

85 Some Facts: Let X be a semimartingale, with X t = M t + A t, where (M t ) is a locally square integrable martingale and (A t ) is a process with bounded variation paths. Let ( X ) t = X t X t, (jump at t) [X,X ] t = ( X ) 2 t, ( X ) 2 s [X,X ]t. <s t If A has bounded variation paths then [A,A]t = ( A) 2 s. <s t Introduction to Stochastic Calculus - 85

86 Moreover, if X is a semimartingale and A is a continuous process with bounded variation paths, then [X,A] t =. To see this t ) X s da s = X t n m t( At n m+1 t A t n m t m= and t X s da s = ) X t n m+1 t( At n m+1 t A t n m t m= as the integral t X s da s is the Riemann-Steltjes and thus [X,A] t = As a result ( )( ) Xt n m+1 t X t n m t At n m+1 t A t n m t =. m= [A,A] t =. Introduction to Stochastic Calculus - 86

87 For a continuous semimartingale (X t ) the decomposition X t = M t + A t into a continuous local martingale (M t ) with M = and a continuous process A with bounded variation paths is unique. The uniqueness is proven as follows. If X t = M t + A t = N t + B t with M,N continuous local martingales, M = N = and A, B processes with bounded variation, then U t = M t N t = B t A t is a continuous local martingale with U =, and has bounded variation paths. Thus, [U,U] t =. Introduction to Stochastic Calculus - 87

88 As a result (recall integration by parts formula) U 2 t = t U s du s and hence Ut 2 is a continuous local martingale with U 2 =. If {τ n } is a localizing sequence, then Ut τ 2 n is a martingale and so E[Ut τ 2 n] =, n and as a result U 2 t τ n = for all t,n. This proves (X t) the decomposition X t = M t + A t into a continuous local martingale (M t ) with M = and a continuous process A with bounded variation paths is unique. Introduction to Stochastic Calculus - 88

89 If X is a semimartingale and f is a locally bounded predictable process, then V t = t f s dx s is also a semimartingale with decomposition V t = N t + B t where X t = M t + A t is a decomposition of X, with M a locally square integrable martingale, A a process with bounded variation paths, and N t = B t = Further, t t f s dn s is a locally square integrable martingale f s da s is a process with bounded variation paths. [V,V ] t = t f 2 s d[x,x ] s. Introduction to Stochastic Calculus - 89

90 Ito formula or Change of variable formula Let f be a smooth function on R and let (G t ) be a continuous function with bounded variation paths. If G were continuously differentiable with derivate g, then h(t) = f (G(t)) is differentiable with continuous derivative f (G(t))g(t) and hence t f (G(t)) =f (G()) + f (G(s))g(s)ds =h() + t f (G(s))dG(s) Introduction to Stochastic Calculus - 9

91 Even if G is not differentiable, one still has t f (G(t)) = f (G()) + f (G(s))dG(s). Assume that f is bounded. The proof uses taylor s expansion: f (b) = f (a) + f (a)(b a) f (c)(b a) 2 where c belongs to the interval {a,b}, i.e. c (a,b) or c (b,a) according as a < b or b < a Introduction to Stochastic Calculus - 91

92 Another fact that is needed is: If G is a continuous function with bounded variation: then there exists a function H(t) such that for any partition for = t < t 1 <... < t m = T of [,T ], m 1 ( ) m i= G(t i+1) G(t i ) H(T ) <, then with tm n = m n ( 2 G(tm n t) G(tm 1 n ( t)) H(t)sup G(t n m t) G(tm 1 n t) ) m m as n as G being continuous on [,t] is uniformly continuous on [,t] and hence sup m ( G(t n m t) G(t n m 1 t) ). Introduction to Stochastic Calculus - 92

93 Observe that for every n with t n m = m n f (G(t)) f (G()) = = m=1 m= one has ( f (G(t n m t)) f (G(t n m 1 t)) ) f (G(t n m 1 t)) ( G(t n m t) G(t n m 1 t) ) m=1 f (θ n,m ) ( G(t n m t) G(t n m 1 t) ) 2 Here θ n,m belongs to the interval {G(t n m 1 t),g(tn m t)}. Since G is a continuous function with bounded variation, the second term goes to zero and first term goes to t f (G(s))dG(s). Introduction to Stochastic Calculus - 93

94 Ito s formula Let X be a continuous semimartingle and let f be a thrice continuously differentiable function on R with f bounded. Then f (X t ) = f (X ) + t f (X s )dx s t f (X s )d[x,x ] s This time we need Taylor s expansion with two terms: f (b) = f (a) + f (a)(b a) f (a)(b a) f (c)(b a) 3 where c belongs to the interval {a,b}, i.e. c (a,b) or c (b,a) according as a < b or b < a. Introduction to Stochastic Calculus - 94

95 We will also need that since X has finite quadratic variation, with tm n = m n one has m X t n m t X t n m 1 t 3 m ( 2 X t n m t X t n m 1 t) sup [X,X ] t sup m as n m ( X t n m t X t n m 1 t ( ) X t n m t X t n m 1 t as uniformly continuity of paths s X) s (ω) of X on [,t] implies sup m ( X t n m t(ω) X t n m 1 t(ω) converges to for every ω. ) Introduction to Stochastic Calculus - 95

96 Recall: we have seen that for every continuous adapted process Y Also that, lim n m= Y t n m t( Xt n m+1 t X t n m t ) t = Yu dx u (11) ( ) 2 Xt n m+1 t X t n m t [X,X ]t. (12) m= On similar lines it can be shown that lim n m= Y t n m t( Xt n m+1 t X t n m t The relation (13) is a weighted version of (12). ) t 2 = Yu d[x,x ] u. (13) Introduction to Stochastic Calculus - 96

97 As in deterministic case, let us write f (X t ) f (X ) = f (X t n m t) ( X t n m+1 t X t n m t m f (X t n m t) ( ) 2 X t n m+1 t X t n m t m f (θ n,m ) ( ) 3 X t n m+1 t X t n m t As noted earlier, the first term converges to t f (X s )dx s and the second term to 2 1 t f (X s )d[x,x ] s. Using boundedness of f and the fact (proven earlier) that Xt n m+1 t X t n m t 3 goes to zero, it follows that third term goes to zero. This completes the proof. ) Introduction to Stochastic Calculus - 97

98 The Ito formula is true even when f is just twice continuously differentiable function. For the proof, one just uses a version of Taylor s theorem with a different remainder form: f (b) =f (a) + f (a)(b a) f (a)(b a) (f (c) f (a))(b a) 2 where c belongs to the interval {a,b}, i.e. c (a,b) or c (b,a) according as a < b or b < a. Introduction to Stochastic Calculus - 98

99 Ito s formula for Brownian motion Let (β t ) be a standard Brownian motion. Then we have noted that [β,β] t = t and hence for a twice continuously differentiable function f, one has t f (β t ) =f (β ) t f (β s )dβ s f (β s )ds Taking f (x) = exp(x) we get, writing Y t = exp(β t ) t Y t = 1 + Y s dβ s + 1 t Y s ds. 2 The last term is the Ito correction term. Introduction to Stochastic Calculus - 99

100 The same proof as outlined earlier, yields for a continuous semimartingale X the second version of Ito formula. Here one uses, for an r.c.l.l. process Z, and semimartingales X,Y t Z u d[x,y ] u n m= = lim ) ) Z t n m t( Xt n m+1 t X t n m t (Yt n m+1 t Y t n m t Introduction to Stochastic Calculus - 1

101 Ito s formula-2 Let X be a continuous semimartingle and let f function on [, ) R such that f t = t f, f x = x f and f xx = 2 2 x f exist and are continuous. Then t t f (t,x t ) =f (,X ) + f t (s,x s )ds + f x (s,x s )dx s t f xx (s,x s )d[x,x ] s Introduction to Stochastic Calculus - 11

102 Let (β t ) be a Brownian motion. Let f (t,x) = exp(σx 1 2 σ 2 t). Let Y t = f (t,β t ). Then noting that f t = t f = 1 2 σ 2 f and f xx = 2 2 x f = σ 2 f, one has Y t = Y + σ In particular, (Y t ) is a martingale. t Y s dβ s. Introduction to Stochastic Calculus - 12

103 If (X t ) takes values in an open convex set U and f is a twice continuously differentiable function, then the same version of Ito formula is valid. In particular, if (X t ) is a (, ) valued continuous semimartingale, then taking f (x) = log(x) we get t log(x t ) = log(x ) + Xs 1 dx s 1 t Xs 2 d[x,x ] s. 2 Once again the last term is the Ito correction term. Introduction to Stochastic Calculus - 13

104 Let us write Y t = t Xs 1 dx s. Then Hence, recalling, we get [Y,Y ] t = t X 2 s d[x,x ] s. t log(x t ) = log(x ) + Xs 1 dx s 1 t 2 log(x t ) = log(x ) + Y t 1 2 [Y,Y ] t Xs 2 d[x,x ] s and hence X t = X exp ( Y t 1 2 [Y,Y ] ) t Introduction to Stochastic Calculus - 14

105 Further, if X is a (, ) valued continuous local martingale, then t M t = Xs 1 dx s is also a local martingale and then X admits a representation X t = X exp ( M t 1 2 [M,M] ) t So every positive continuous local martingale is of the form given above for a local martingale M. Introduction to Stochastic Calculus - 15

106 Ito s formula-3 Let X be a continuous semimartingle, A be a continuous process with bounded variation and let f = f (a,x) function on R R such that f a = a f, f x = x f and f xx = 2 2 x f exist and are continuous. Then t t f (A t,x t ) =f (A,X ) + f a (A s,x s )das + f x (A s,x s )dx s t f xx (A s,x s )d[x,x ] s Introduction to Stochastic Calculus - 16

107 In typical applications of Ito s formula, one needs to show a certain process is a martingale. First one shows that it is a local martingale by expressing it as an integral w.r.t. a locally square integrable martingale and then uses other methods to show it is a martingale. Let (M t ) be a continuous locally square integrable martingale and let (h s ) be a locally bounded predictable process. Question: for what values of λ is Z a martingale where { t t Z t = exp σ h s dm s + λ hs 2 d[m,m] s }. Introduction to Stochastic Calculus - 17

108 Let Y t = t h s dm s and V t = t h 2 s d[m,m] s. Note that (Y t ) is a square integrable local martingale and V t = [Y,Y ] t. Let f (a,x) = exp(σx + λa). Then Z t = f (V t,y t ). Now noting that f a = a f = λf and f xx = 2 2 x f = σ 2 f, one has Z t = Z + σ t Z s dy s + (λ + 1 t 2 σ 2 ) Z s dv s. Introduction to Stochastic Calculus - 18

109 Hence if λ = 1 2 σ 2, it follows that Z t = Z + σ t Z s dy s. Hence for λ = 1 2 σ 2, Z is a local martingale. It can be shown that this is a necessary and sufficient condition. Of course, we will need some other way to prove that the local martingale is a martingale. Introduction to Stochastic Calculus - 19

110 Levy s characterization of Brownian motion. Suppose M is a continuous local martingale with M = and such that Z t = (M t ) 2 t is also a local martingale, or [M,M] t = t, then M is a brownian motion. Apply Ito formula to (for a fixed real number λ) { f (t,x) = exp iλx + λ 2 t 2 We note that 1 2 f xx = f t and hence Z t = exp { iλm t + λ 2 } t 2 is a local martingale. Since Z is bounded, it is a martingale. }. Introduction to Stochastic Calculus - 11

111 As a result, E[Z t F s ] = Z s i.e. [ E exp { iλm t + λ 2 t } ] Fs = exp { iλm s + λ 2 s 2 }. 2 2 Hence [ E exp { iλ(m t M s ) + λ 2 (t s) } ] Fs = 1. 2 or E [exp { iλ(m t M s ) } ] F s = exp { λ 2 (t s) }. 2 Thus [ E exp { iλ(m t M s ) }] = exp { λ 2 (t s) } 2 and so (M t M s ) has normal distribution, mean Variance t s. Introduction to Stochastic Calculus - 111

112 Further, E [exp { iλ(m t M s ) } ] F s = E [exp { iλ(m t M s ) }] implies that (M t M s ) is independent of any set of F s measurable random variables. In particular, (M t M s ) is independent of M u1,m u2,...,m uk. Thus M t is a Brownian motion. Introduction to Stochastic Calculus - 112

113 Sometimes, given a process (Z t ) one needs to find a process (A t ) with bounded variation paths such that M t = Z t A t is a local martingale or a martingale (U t ) such that N t = Z t U t is a local martingale. Example: Let (β t ) be the standard Brownian motion and for σ > and µ R let S t = S exp{σβ t + µt}. Introduction to Stochastic Calculus - 113

114 We need to find a local martingale (U t ) such that Z t = S t U t is a martingale. Let us try U t = exp{λβ t 1 2 λ 2 t}. We know this is a martingale. Now { Z t = S exp (σ + λ)β t 1 } 2 (λ 2 2µ) Thus (Z t ) would be a martingale if (σ + λ) 2 = (λ 2 2µ). Thus σ 2 + 2σλ = 2µ and hence. λ = µ+ 1 2 σ 2 σ In this case, (U t ) and (Z t ) turn out to be martingales. Introduction to Stochastic Calculus - 114

115 Ito s formula-4 Let X 1,X 2,...,X d be a continuous semimartingles and let f function on [, ) R d such that f t = t f, f i = x i f and f ij = 2 x i x j f exist and are continuous. Let X t = (X 1 t,...,x d t ). Then t f (t,x t ) =f (,X ) f t (s,x s )ds + d t i=1 d t f ij (s,x s )d[x i,x j ] s i,j=1 f i (s,x s )dx i s Introduction to Stochastic Calculus - 115

116 Often one needs to compute the part with bounded variation paths separately. In light of this, we can recast the formula showing the (local) martingale part and the part with bounded variation paths separately. Let X 1,...,X d be continuous semimartingales with X i = M i + A i where M i are continuous local martingales and A i be processes with bounded variation paths. Let f C 1,2 ([, ) R d b ). Then Introduction to Stochastic Calculus - 116

117 where and N t = t B t =f (,X ) d j=1 d k=1 f (t,x t ) = N t + B t d t f j (s,x s )dms j j=1 f t (s,x s )ds + d j=1 t t f jk (s,x s )d[m j,m k ] s f j (s,x s )da j s Introduction to Stochastic Calculus - 117

118 Ito s formula for r.c.l.l. semimartingales. Let X 1,...,X d be semimartingales, X t := (X 1 t,...,x d t ). Let f C 1,2 ([, ) R d ). t f (t,x t ) =f (,X ) + + d j=1 t d j=1 d k=1 f t (s,x s )ds f j (s,x s )dx j s t f jk (s,x s )d[x j,x k ] + <s t {f (s,x s ) f (s,x s ) d j=1 f j(s,x s ) Xs j d j=1 d 1 } k=1 2 f jk(s,x s )( Xs j )( Xs k ). Introduction to Stochastic Calculus - 118

119 Martingale representation theorem Let (β t ) be a standard Brownian motion. Let F t = σ{β u : u t} Let (M t ) be a martingale w.r.t. (F t ). Then M admits a representation t M t = M + h s dβ s, t T for some predictable process (h s ) such that T h 2 s ds < a.s. Introduction to Stochastic Calculus - 119

120 For T >, let Z be any random variable such that E[ Z ] <. Then taking the martingale M t = E[Z F t ], we can get predictable process (h s ) such that T h 2 s ds < a.s. and t Z = M T = M + h s dβ s. Introduction to Stochastic Calculus - 12

121 Martingale representation theorem- multidimensional Let (β 1 t,...,β d t ) be d- independent standard Brownian motions. Let F t = σ{β j u : u t,1 j d} Let (M t ) be a martingale w.r.t. (F t ). Then M admits a representation M t = M + d t j=1 h j sdβ j s, t T for some predictable process (h s ) such that T (h j s) 2 ds < a.s., 1 j d. Introduction to Stochastic Calculus - 121

122 Suppose X is a continuous process on (Ω,F,P) and let Q be an equivalent probability measure on (Ω,F ) such that X is a martingale on (Ω,F,Q). Equivalent means P(A) = if and only if Q(A) =. Such a Q is called an Equivalent Martingale Measure, written as EMM, for X and is of great importance in Mathematical finance. Introduction to Stochastic Calculus - 122

123 Let R denote the Radon-Nikodym derivative of Q w.r.t. P, em i.e. Q(A) = RdP. Equivalence of P and Q means P(R > ) = 1. Let R t = E P [R F t ]. Then as noted earlier (R t ) is a martingale. Moreover, for A F t Q(A) = RdP A = E[R F t ]dp. A = R t dp Thus R t is the R-N derivative of Q w.r.t. P on F t. A A Introduction to Stochastic Calculus - 123

124 Now, (X t ) being a martingale on (Ω,F,Q) means A X t dq = A X s dq for s < t and a F s. But this is same as X t R t dp = X s R s dp A since R u is the R-N derivative of Q w.r.t. P on F u. A Introduction to Stochastic Calculus - 124

125 Hence, (X t R t ) is a martingale on (Ω,F,P). Further, R t being a positive martingale implies Rt 1 is a semimartingale and thus X t = X t R t.r 1 t is a semimartingale. Suppose X t = N t + B t is the decomposition of X with local martingale N. We have seen that where M t = t R 1 s dr s. R t = R exp(m t 1 2 [M,M] t) Introduction to Stochastic Calculus - 125

126 Hence Z t = (N t + B t )exp(m t 1 2 [M,M] t) is a martingale. So we see that if an EMM exists for X, then X must be a semimartingale and further the R-N derivative for the EMM w.r.t. the model probability measure on F t is R t = R exp(m t 1 2 [M,M] t) where M is a local martingale such that is a martingale. Z t = (N t + B t )exp(m t 1 2 [M,M] t) Introduction to Stochastic Calculus - 126