Survey

# Download Comp_Phys-ito

Document related concepts

Transcript

E. Vitali Lecture 7. Ito integral In this chapter we will introduce the basic notions of stochastic calculus, starting from brownian motion. When dealing with stochastic processes, describing phenomena depending on time, one wishes to have tools to study the functions of stochastic processes, typically performing “derivatives” or “integrals”. Stochastic calculus is the branch of mathematics dealing with this important topic. The reason why traditional calculus is not suitable for stochastic processes, relying on the brownian motion, lies essentially in the basic fact √ that V ar(Bt ) = t, implying that Bt “scales” as t, and thus has non differentiable trajectories. In order to properly fix the mathematical environment, we assign once and for all a stochastic basis in usual ipothesis: Ω, F , {Ft }t≥0 , P (1) where we suppose a continuous brownian motion is defined, with increments independent of the past. For simplicity, we consider now only the one-dimensional case: B = Ω, F , {Ft }t≥0 , {Bt }t≥0 , P On the other hand, if we take the average: (2) Nota 1 Having fixed the stochastic basis, we will use the simple notation X = {Xt }t≥0 to indicate processes. Naturally, we will be careful to verify that all processes are adapted, i.e. Xt is Ft -measurable. We start from the observation that a stochastic process X = {Xt }t≥0 can be viewed as a function of two variables: X : [0, +∞] × Ω → R, (t, ω) Xt (ω) (3) where ω specifies the random trajectory and t is the time instant. We will learn in this chapter to define two kinds of integration of processes, one with respect to time, and the other with respect to brownian motion: Z β Z β Xs ds, Xs dBs (4) α α and to give a precise meaning to expressions of the form: dXt = Ft dt + Gt dBt (5) which will turn out to be a very important tool to build up new processes starting from brownian motion and will be the starting point of the theory of stochastic differential equation which we will develop in the next chapter. In this chapter we will provide rigorous definitions and results about stochastic calculus, together with examples and applications. We will omit the proofs of several theorems, which would require more advanced tools and can be found in excellent textbooks on this subject. 1 E. Vitali I. Lecture 7. Ito integral INTEGRATION OF PROCESSES WITH RESPECT TO TIME We start thus from a process, interpreted as a function of two variables: X : [0, +∞) × Ω → R, (t, ω) Xt (ω) (6) From the very definition of a stochastic process, we know that, for fixed t, the function: ω Xt (ω) is F -measurable and in particular Ft measurable. On the other hand, we don’t know, a priori, for fixed ω, measurability properties of the function t Xt (ω), a real valued function defined on [0, +∞). We will focus our attention to progressively measurable processes, that is, by definition, processes such that, for all t > 0, the function (t, ω) Xt (ω) is B([0, t ]) ⊗ Ft -measurable. It is possible to show that such technical ipothesis surely holds if the given process is continuous, i.e. it has continuous trajectories. For such processes, the function: Xt (ω) t (7) is measurable and we can define time integrals, one for each trajectory: Z β Z β def Xs ds (ω) = Xs (ω)ds (8) α α where we have fixed a time interval [α, β], 0 ≤ α < β < +∞. Thanks to the ipothesis of progressive measurability, the integral, if it exists, is a random variable. It is thus natural to introduce the following: Definizione 2 Let Λ1 (α, β) be the set made of equivalence classes of progressively measurable processes such that: Z β P |Xs |ds < +∞ = 1 (9) α where we consider equivalent two processes X and Y such that: Z β P |Xt − Yt |dt = 0 = 1 (10) α As usual, we often neglect the difference between a process and an equivalence class of processes. Since we have chosen to work under usual ipothesis, if X ∈ Λ1 (α, β) we can always turn to a modification such that the integral: Z β Xs ds (11) α exists for any ω. Such integral defines a real random variable Fβ measurable and, if: Z β E [|Xs |] ds < +∞ (12) α 2 E. Vitali Lecture 7. Ito integral then, by classical Fubini’s theorem: Z β Z E Xs ds = α β E [Xs ] ds (13) α Moreover, if X ∈ Λ1 (0, T ), then the function: Z t (t, ω) Xs ds, t ∈ [0, T ] (14) 0 defines a stochastic process: Z t Xs ds 0 (15) t which is continuous, since any integral is continuous with respect to the extremum, and thus progressively measurable. To summarize, a stochastic process, under some quite natural ipothesis, can be integrated respect to time: this is a simple Lebesgue integral of the single trajectories. II. ITO STOCHASTIC INTEGRAL We are now going to build up a quite different integration, with respect to the brownian motion. A. Stochastic integral of elementary processes As before, we fix a time interval [α, β], 0 ≤ α < β < +∞ and we start with the following: Definizione 3 We say that a stochastic process X is elementary if: Xt (ω) = n−1 X ei (ω)1[ti ,ti+1 ) (t) + en (ω)1{β} (t) (16) i=0 for some choice of the integer n and of the times α = t0 < · · · < tn = β. ei are random variables Fti -measurable. An elementary process remains equal to a random constant over finite intervals of time. By construction, such a process is progressively measurable. We give now the first basic: Definizione 4 If X is anRelementary process, we call stochastic inteβ gral of X and we denote α Xs dBs the real random variable: Z β α def Xs dBs (ω) = n−1 X i=0 ei (ω) Bti+1 (ω) − Bti (ω) 3 (17) E. Vitali Lecture 7. Ito integral We denote S(α, β) the set of elementary processes, and S 2 (α, β) the set of square-integrable elementary processes, i.e. E [|Xt |2 ] < +∞. Naturally, X ∈ S 2 (α, β) if and only if E[e2i ] < +∞. Le’s study now the map: Z β def S(α, β) ∋ X I(X) = Xs dBs (18) α Le’s enumerate some important properties: 1. it is linear; 2. since brownian motion is adapted, I(X) is Fβ -measurable because it depends only of random variables at times preceding β; 3. it satisfies the following additivity property: Z β Z γ Z β Xs dBs = Xs dBs + Xs dBs , α < γ < β α α (19) γ 4. If X ∈ S 2 (α, β), then I(X) is square integrable: E "Z β Xs dBs α 2 # < +∞ (20) To verify this property we show that ei ej Bti+1 − Bti Btj+1 − Bti is integrable: : if i = j, we know that e2i is integrable since 2 X ∈ S 2 (α, β); Bti+1 − Bti is naturally integrable (it is a brownian motion!) and independent of e2i which is Fti -measurable. Thus 2 e2i Bti+1 − Bti is integrable being the product of integrable independent random variables. This implies that ei Bti+1 − Bti is square-integrable, and thus ei ej Bti+1 − Bti Btj+1 − Bti is integrable being a product of square-integrable random variables. We have thus defined a linear map: 2 2 I : S (α, β) → L (Ω, Fβ , P ), X def I(X) = Z β Xs dBs (21) α Since I(X) is square-integrable, it is also integrable because the probability P is a finite measure; it thus does make sense to perform the following calculation: E hR β i Pn−1 X dB |F s s α = E i=0 ei Bti+1 − Bti |Fα = α Pn−1 = i=0 E E ei Bti+1 − Bti |Fti |Fα = Pn−1 E ei E Bti+1 − Bti |Fti |Fα = = i=0 Pn−1 E ei E Bti+1 − Bti |Fα = 0 = i=0 4 (22) E. Vitali Lecture 7. Ito integral where we have used the fact that ei is Fti -measurable and that Bti+1 − Bti is independent of Fti , together wiith the properties of conditional expectation. Thus we have found that: Z β E Xs dBs |Fα = 0 (23) α which implies, taking the expectation of both members, that: Z β E Xs dBs = 0 (24) α . Let’s now evaluate: Rβ E = α Xs dBs 2 |Fα = (25) h i P 2 2 B − B E e |F Btj+1 − Btj |Fα ti+1 ti α + i i i6=j E ei ej Bti+1 − Bti P If i < j, we have: E ei ej Bti+1 − Bti Btj+1 − Btj |Fα = = E E ei ej Bti+1 − Bti Btj+1 − Btj |Ftj |Fα = = E ei ej Bti+1 − Bti E Btj+1 − Btj |Ftj |Fα = 0 (26) so that no contribution arises from non diagonal terms. We have thus: E R β α Xs dBs 2 |Fα = (27) i 2 P h = i E e2i Bti+1 − Bti |Fα = i i 2 P h h 2 = i E E ei Bti+1 − Bti |Fti |Fα = h i i 2 P h = i E e2i E Bti+1 − Bti |Fti |Fα = h i 2 i P h = i E e2i E Bti+1 − Bti |Fα = hR i P 2 β 2 = E [ i ei (ti+1 − ti ) |Fα ] = E α Xs ds|Fα We conclude that: "Z β E Xs dBs α 2 # |Fα = E Z β Xs2 ds|Fα α which implies the following very important equality: "Z 2 # Z β β 2 E =E Xs dBs Xs ds α α carrying a deep geometrical meaning, as we will see in a moment. 5 (28) (29) E. Vitali Lecture 7. Ito integral The linear map 2 2 I : S (α, β) → L (Ω, Fβ , P ), Z def I(X) = X β (30) Xs dBs α satisfies the following: ||I(X)||2L2(Ω,Fβ ,P ) =E Z β Xs2 ds α (31) The right hand side of this equality is a double integral in dt and P (dω) of the square of the function (t, ω) Xt (ω) (32) so that it has the form of a square norm in a L2 -like space which we define in the following: Definizione 5 We denote M 2 (α, β) the set of equivalence classes of progressively measurable processes such that: Z β 2 E Xs ds < +∞ (33) α where we identify two processes X and Y if: Z β P |Xt − Yt |dt = 0 = 1 (34) α M 2 (α, β) is an Hilbert space, subspace of L2 ((α, β) × Ω, B(α, β) ⊗ Fβ , λ ⊗ P ), where λ is the Lebesgue measure. We have thus built an isometry, called Ito isometry: ||I(X)||2L2(Ω,Fβ ,P ) = ||X||2M 2(α,β) B. (35) First extension of Ito integral The map: 2 2 2 I : S (α, β) ⊂ M (α, β) → L (Ω, Fβ , P ), X def I(X) = Z β α Xs dBs (36) is linear and isometric: ||I(X)||2L2(Ω,Fβ ,P ) = ||X||2M 2(α,β) (37) and thus it is bounded. Moreover, it is possible to show that S 2 (α, β) is a dense subset of M2 (α, β), M 2 (α, β), that is, for each process X ∈ (n) 2 there exists a sequence Y of elementary processes in S (α, β) such n that: 6 E. Vitali Lecture 7. Ito integral X = L2 − lim Y (n) (38) n→+∞ where the notation L2 − lim means that: Z β n→+∞ (n) 2 E |Ys − Xs | ds → 0 (39) α The bounded extension theorem from functional analysis guarantees the possibility of giving the following: Definizione 6 We call stochastic integral of a process X ∈ M 2 (α, β), Rβ and we denote α Xs dBs , the element of L2 (Ω, Fβ , P ): Z β Z β def Xs dBs = lim Ys(n) dBs (40) n→∞ α α where Y (n) n ⊂ S 2 (α, β) is any sequence of elementary squareintegrable processes converging to X in M 2 (α, β). The above written limit is meant in the geometry of L2 (Ω, Fβ , P ). The map I becomes thus an isometry between Hilbert spaces: 2 2 I : M (α, β) → L (Ω, Fβ , P ), X def I(X) = Z β Xs dBs (41) α The properties of the restriction to S 2 (α, β) guarantee that all the properties of I(X) discussed above, including the ones about expectations and conditional expectations, hold for each X ∈ M 2 (α, β). C. Second extension of Ito integral It is possible to extend the definition of the stochastic integral to processes not belonging to M 2 (α, β). We won’t enter the details of this extension, we just sketch the procedure which relies on approximating sequences. Definizione 7 We let Λ2 (α, β) be the set of equivalence classes of progressively measurable processes such that: Z β 2 P |Xs | ds < +∞ = 1 (42) α where we identify two processes X and Y if: Z β P |Xt − Yt |dt = 0 = 1 α 7 (43) E. Vitali Lecture 7. Ito integral Naturally M 2 (α, β) ⊂ Λ2 (α, β). It is possible to show that for each X ∈ Λ2 (α, β) there exists process (n) a sequence of elementary processes Y in Λ2 (α, β), such that: n Z β |Ys(n) − Xs |2 ds → 0, n → +∞ (44) α where the limit is meant in probability. Then it turns out that the sequence: Z β (n) Ys dBs (45) α n converges in probability to a random variable which depends on X but not on the approximating sequence. Such random variable provides the natural definition of Ito integral of the process X: Z β Z β def Xs dBs = P − lim Ys(n) dBs (46) n→+∞ α α 2 It is quite simple to show that, if X ∈ M (α, β) this definition coincides to the one given above. When dealing with processes not belonging to M 2 (α, β) care has to be taken since properties involving expectations and conditional expectations are no longer valid. Finally, we mention without proof a quite intuitive result, stating that, for continuous processes, Riemann sums: n−1 X i=0 Xti (Bti+1 − Bti ) (47) converge in probability to Ito integral as the width of the partition tends to 0. D. The Ito integral as a function of time A crucial object for the study of stochastic calculus is the following: Z t def I(t) = Xs dBs (48) 0 2 where X ∈ Λ (0, T ), and the instant t belongs to the interval [0, T ]. The following very simple but important property holds for s < t: Z t I(t) = I(s) + Xs dBs (49) s Moreover, I(t) is Ft -measurable for all t; we already know that thisis the case if X is an elementary process. In the general case, if Y (n) n is a sequence of elementary processes nR o approximating in probability X in t (n) 2 Λ (0, T ), and In (t) = 0 Ys dBs , then In (t) is Ft -measurable. Since n In (t) converges to I(t) in probability, it is possible to extract a subsequence converging almost surely to I(t), which is thus Ft -measurable. 8 E. Vitali Lecture 7. Ito integral It is possible to show, maybe turning to a modification, that I(t) is continuous. In the particular case X ∈ M 2 (0, T ), we already know that I(t) is square-integrable and: Z t E [I(t)|Fs ] = I(s) + E Xs dBs |Fs = I(s) (50) s so that I(t) is a square-integrable martingale. E. Wiener integral A very important special case happens when the integrand is non random. Let f : [0, T ] → R be a square-integrable real valued function f ∈ L2 (0, T ). The process: (t, ω) f (t) (51) independent of ω, belongs to M 2 (0, T ). In such case: Z t I(t) = f (s)dBs (52) 0 is called Wiener integral of f on [0, t]. We know that: Z t 2 E[I(t)] = 0, E[I(t) ] = f 2 (s)ds (53) 0 If f is piece-wise constant: f (s) = n−1 X ci 1[ti ,ti+1 ) (s) (54) i=0 then: I(t) = n−1 X i=0;ti <t,ti+1 <t ci Bti+1 − Bti (55) and thus I(t) is normal, being a linear combination of normal random variables. This holds also for (I(t1 ), . . . , I(tn )), which turns out to be normal. This property continues to hold in the limit in M 2 (0, T ), since the convergence in the sense of L2 preserves the normal character of laws of random variables. We conclude that: Z t Z t 2 I(t) = f (s)dBs ∼ N 0, f (s)ds (56) 0 III. 0 STOCHASTIC DIFFERENTIAL Till now we have learned to define integrals of stochastic properties, both with respect to time and with respect to brownian motion. Before 9 E. Vitali Lecture 7. Ito integral proceeding, we briefly summarize what we have done. We have fixed the mathematical environment assigning a stochastic basis: Ω, F , {Ft }t≥0 , P (57) in usual ipothesis, that is with a right-continuous filtration such that Ft contains all the elements of F whose probability is zero. Moreover, we start from a one-dimensional continuous brownian motion {Bt }t with increments independent of the past. Given a time interval [0, T ], 1. if F ∈ Λ1 (0, R t T ), maybe turning to a modification, we can build up a process 0 Fs ds, for t ∈ [0, T ] continuous, and thus progressively measurable. Moreover the trajectories of such process are integrable and also square-integrable being continuous on the compact interval [0, T ], so that: Z t Fs ds 0 0≤t≤T ∈ Λ1 (0, T ) ∩ Λ2 (0, T ) = Λ2 (0, T ) (58) 2. if G ∈ Λ2 (0, TR), maybe turning to a modification, we can build t up a process 0 Gs dBs , for t ∈ [0, T ] continuous, and thus progressively measurable. Moreover the trajectories of such process are integrable and also square-integrable being continuous on the compact interval [0, T ], so that: Z 0 t Gs dBs 0≤t≤T ∈ Λ1 (0, T ) ∩ Λ2 (0, T ) = Λ2 (0, T ) (59) Definizione 8 Let {Xt }t≥0 be a process such that, ∀t ∈ [0, T ], the following equality holds: Z t Z t Xt = X0 + Fs ds + Gs dBs (60) 0 0 X0 being a F0 -measurable random variable, F ∈ Λ1 (0, T ) and G ∈ Λ2 (0, T ). We say that {Xt }t≥0 is an Ito process or, equivalently, that {Xt }t≥0 has stochastic differential: dXt = Ft dt + Gt dBt (61) An extremely important result, which we will state without proof, is the following: Teorema 9 (Ito formula) Let X (i) , i = 1, . . . , m be a collection of Ito processes with differentials: (i) (i) (i) dXt = Ft dt + Gt dBt (62) with F (i) ∈ Λ1 (0, T ) and G(i) ∈ Λ2 (0, T ). Let also f : Rm × R+ → R be a measurable function, continuous in every point (x, t), x = (x1 , . . . , xm ), continuously differentiable twice in x and once in t. Then, writing 10 E. Vitali def Lecture 7. Ito integral (1) (m) Xt = (Xt , . . . , Xt differential: dYt = ft (Xt , t)dt+ def ), the process Yt = f (Xt , t) is an Ito process with m X 1 (i) fxi (Xt , t)dXt + m X 2 i,j=1 i=1 (i) (j) fxi xj (Xt , t)Gt Gt dt (63) that is: dYt = (64) Pm P (i) (i) (j) = ft (Xt , t) + i=1 fxi (Xt , t)Ft + 12 m dt + i,j=1 fxi xj (Xt , t)Gt Gt P (i) m dBt + i=1 fxi (Xt , t)Gt IV. EXTENSION TO THE MULTIDIMENSIONAL CASE We conclude this chapter sketching the extension of the stochastic calculus formalism to the multidimensional case. The generalization is straightforward: nothing actually change but a slight modification of the notations. The starting point is, as usual, a stochastic basis in usual ipothesis where a continuous d-dimensional Brownian motion with increments independent of the past is assigned once and for all. The processes Ft of the previous paragraphs take now values in Rm while the processes Gt take values in Rm×d and we look at each component. We say that Ft belongs to Λ1m (0, T ), T > 0 if Fi,t belongs to Λ1 (0, T ) for all i = 1, . . . , m. In the same way. we say that Gt be2 longs to Λ2m,d (0, T ) (respectively Mm,d (0, T )) if Gij,t belongs to Λ2 (0, T ) (respectively M 2 (0, T ))R ) for all i = 1, . . . , m, j = 1, . . . d. t The time integral 0 Fs ds is defined as the vector of components Rt Rt F ds, while 0 Gs dBs is defined as the vector of components 0 i,sR P t d j=1 0 Gij,s dBj,s . Ito isometry becomes: "Z 2 # Z t t E |Gs |2 ds E Gs dBs = (65) 0 0 2 Mm,d (0, T ). m and holds whenever Gt ∈ We observe that | | in the left hand side denotes the norm in R , while in the right hand side it denotes the norm in Rm×d . If we can write: Z t Z t Xt = X0 + Fs ds + Gs dBs (66) 0 0 with X0 F0 -measurable we say that X is an Ito process, or, equivalently, that X has stochastic differential: dXt = Ft dt + Gt dBt Ito’s formula can be generalized as follows: 11 (67) E. Vitali Lecture 7. Ito integral Teorema 10 (Multidimensional Ito formula) Let X be a process taking values in Rm with stochastic differential: dXt = Ft dt + Gt dBt (68) with Ft ∈ Λ1m (0, T ) and G ∈ Λ2m,d (0, T ). We let also f : Rm × R+ → R be a measurable function, continuous in every point (x, t), x = (x1 , . . . , xm ), continuously differentiable twice in x and once in t. Then the process def Yt = f (Xt , t) admits stochastic differential: dYt = ft (Xt , t)dt + m X fxi (Xt , t)dXi,t + i=1 m d X 1X fxi xj (Xt , t) Gih,t Gjh,tdt 2 i,j=1 h=1 (69) 12