Brownian Motion March 15, 2013 Contents 1 The Gaussian Process 2 2 Markov Processes 14 3 Markov Transition Functions 17 4 Strong Markov Property 19 5 Brownian Martingales 23 6 Donsker’s Theorem 33 1 1 The Gaussian Process Definition 1.1. Brownian Motion A Brownian motion is a stochastic process {Bt }t≥0 on R s.t. • t → Bt is continuous a.s. N • For positive strictly increasing sequence {ti }N i=1 we have that {Bti − Bti−1 }i=2 are independent. • For 0 ≤ s < t we have that Bt−s ∼ N (0, t − s) Definition 1.2. d-Dimensional Brownian Motion (1) (d) (i) {Bt }t≥0 : Bt = (Bt , ..., Bt ) is a d-dimensional Brownian motion if each Bt is an independent Brownain motion on R. Definition 1.3. Characteristic Function For a random variable Z the characteristic function is ϕZ (θ) := E[eiθZ ] Definition 1.4. Gaussian Vector Pd Z = (Z1 , ..., Zd ) ∈ Rd is a Gaussian vector if any linear combination i=1 λi Zi is a Gaussian random variable in R i.e. has density (x−µ)2 2σ 2 e− f (x) = √ 2πσ Corollary 1.1. The characteristic function of a Gaussian random variable N (µ, σ 2 ) is ϕZ (θ) = eiθµ−σ 2 2 θ /2 Proof. Let Z ∼ N (µ, σ 2 ) ϕZ (θ) = E[eiθz ] (z−µ)2 Z ∞ − 2σ2 iθz e √ = e dz 2πσ −∞ Z ∞ (z−µ)2 −2σ 2 iθz 1 2σ 2 =√ e− dz 2πσ −∞ Z ∞ z 2 −2µz+µ2 −2σ 2 iθz 1 2σ 2 =√ e− dz 2πσ −∞ Z ∞ (z−(µ+σ 2 iθ))2 +σ 4 θ 2 −2µσ 2 iθ 1 2σ 2 =√ e− dz 2πσ −∞ Z ∞ − (z−(µ+σ22 iθ))2 2σ e −σ 2 θ 2 /2+µiθ √ =e dz 2πσ −∞ = e−σ 2 2 θ /2+µiθ Definition 1.5. Gaussian Process The stochastic process {Xt }t≥0 is Gaussian if (Xt1 , ..., XtN ) is a Gaussian vector for any choice of {ti }N i=1 Lemma 1.1. The following are facts about Gaussian random variables: • Z ∼ N (µ, σ 2 ), c > 0 =⇒ cZ ∼ N (cµ, c2 σ 2 ) • Zi ∼ N (µi , σi2 ) i = 1, 2 independent then Z1 + Z2 ∼ N (µ1 + µ2 , σ12 + σ22 ) 2 • If Zk ∼ N (µk , σk2 ) for k ∈ N then limk→∞ µk = µ, limk→∞ σk2 = σ 2 and Zk converges a.s. to some Z ∼ N (µ, σ 2 ) • ϕZ (θ) determines the law of Z ` • Z1 Z2 ⇐⇒ E[eiθ1 Z1 eiθ2 Z2 ] = E[eiθ1 Z1 ]E[eiθ2 Z2 ] ∀θ1 , θ2 • If limk→∞ ϕZk (θ) = ϕZ (θ) then Zk converges to Z in distribution. • The law of the Gaussian vector Z = (Z1 , ..., ZN ) is determined by E[Zk ], E[Zi Zj ] • If Z1 , Z2 are Gaussian then they are independent iff E[Z1 Z2 ] = E[Z1 ]E[Z2 ] Lemma 1.2. A Brownian motion is a Gaussian process PN Proof. We need to check that i=1 λi Bti is Gaussian for any choice of {λi , ti }N i=1 . We can write N N X X λ i B ti = µi (Bti − Bti−1 ) i=1 i=1 for some µ where t0 = 0, B0 = 0. This form is the sum of independent Gaussian random variables and hence is itself Gaussian. Lemma 1.3. The Brownian Bridge: Xt := Bt − tB1 for t ∈ [0, 1] is a Gaussian process PN Proof. We need to check that for any choice of {λi , ti }N i=1 we have that i=1 λi (Bti − ti B1 ) is Gaussian. This sum can be written as ! N N n N +1 X X X X λi (Bti − ti B1 ) = λi Bti − λi ti B1 = λ̂i Bti i=1 i=1 i=1 i=1 Pn where λ̂i = λi for i ≤ N , λ̂N +1 = − ( i=1 λi ti ) and tN +1 = 1 This is the same summation form as the Brownian motion and hence is Gaussian. Lemma 1.4. The Orstein-Uhlenbeck process: Xt := e−Ct Be2Ct for C > 0 is a Gaussian process PN −Cti Proof. We need to check that for any choice of {λi , ti }N Be2Cti is i=1 we have that i=1 λi e Gaussian. Write λ̂i := λi e−Cti and then t̂i = e2Cti . We then have that the sum is of the form: N X λi e−Cti Be2Cti = i=1 N X λ̂i Bt̂i i=1 then using the same rearrangment trick as in the Brownian motion case we have the desired result. Lemma 1.5. For Brownian motion Bt we have that E[Bt Bs ] = min(t, s) Proof. WLOG let 0 ≤ s < t, then we have that E[Bt Bs ] = E[(Bt − Bs )Bs + Bs2 ] = E[(Bt − Bs )(Bs − B0 ) + (Bs − B0 )2 ] = E[Bt − Bs ]E[Bs − B0 ] + E[(Bs − B0 )2 ] by independence =0+s 3 Theorem 1.1. If X = {Xt }t≥0 is a Gaussian process with continuous paths s.t. • E[Xt ] = 0 ∀t • E[Xt Xs ] = min(s, t) then X is a Brownian motion. Proof. We already have that the paths are continuous a.s. and the expectation is zero so it remains to show that the increments Xt − Xs are independent Gaussians with zero mean and variance t − s. By definition of a Gaussian process Xt − Xs is Gaussian. Let 0 ≤ s < t E[Xt − Xs ] = E[Xt ] − E[Xs ] =0−0 =0 V ar[Xt − Xs ] = E[(Xt − Xs )2 ] − E[Xt − Xs ]2 = E[Xt2 + Xs2 − 2Xt Xs ] = E[Xt2 ] + E[Xs2 ] − 2E[Xt Xs ] = t + s − 2min{t, s} =t−s We now need to show that the increments are independent. It suffices to show that Xt1 − Xt2 , Xt3 − Xt4 are independent for 0 ≤ t4 < t3 ≤ t2 < t1 E[Xt1 − Xt2 ]E[Xt3 − Xt4 ] = 0 E[(Xt1 − Xt2 )(Xt3 − Xt4 )] = E[Xt1 Xt3 − Xt1 Xt4 − Xt2 Xt3 + Xt2 Xt4 ] = E[Xt1 Xt3 ] − E[Xt1 Xt4 ] − E[Xt2 Xt3 ] + E[Xt2 Xt4 ] = t3 − t4 − t3 + t4 =0 which suffices for independence since the increments are Gaussian. Lemma 1.6. If Bt is a Brownian motion then 1 Z I= Bt dt 0 is a Gaussian N (0, 1/3) variable. Proof. I is the a.s. limit as N → ∞ of N N 1 X 1 X Bk/N = (N − k + 1)(B k − B k−1 ) N N N N i=1 k=1 The terms in the second sum are independent Gaussian random variables and hence the sum is a 4 Gaussian random variable moreover the almost sure limit is Gaussian. Z 1 Bt dt E[I] = E 0 Z 1 E[Bt ]dt = 0 =0 V ar[I] = E[I 2 ] Z 1 Z Bs ds =E 0 Z 1 Z 1 Bt dt 0 1 E[Xt Xs ]dsdt = Z by Fubini 0 0 1 Z 1 min{t, s}dsdt = 0 0 Z 1Z t 1 Z 1 Z tdsdt sdsdt + = 0 0 Z 1 = 0 t2 /2dt + t 0 1 Z t − t2 dt 0 = 1/6 + 1/2 − 1/3 = 1/3 So providing the use of Fubini is valid we have the desired result. We need that Z 1 Z 1 E |Bt ||Bs |dsdt < ∞ 0 We have that: Z 1 Z 1 Z E |Bt ||Bs |dsdt = 0 0 1 Z 1 E[|Bt ||Bs |]dsdt 0 Z 0 since psoitive 0 1 Z 1 ≤ 0 Z 1 0 Z 1 0 0 q √ E[Bt2 ]E[Bs2 ]dsdt by Cauchy-Schwarz ts <1 Lemma 1.7. Scaling Lemma: If B is a Brownian motion and c > 0 then Xt := 1c Bc2 t for t ≥ 0 is also a Brownian motion. Proof. It suffices to check that X has continuous paths, E[Xt ] = 0, E[Xt Xs ] = min{t, s} and that X is Gaussian. Xt is a continuous function of a Brownian motion hence clearly has continuous paths. E[Xt ] = 1c E[Bc2 t ] = 0 since B is a Brownian motion. 5 Let s < t then: 1 1 E[Xs Xt ] = E Bc2 s Bc2 t c c 1 = 2 E[Bc2 s Bc2 t ] c 1 = 2 c2 s c =s N X λk Xtk = k=1 N X λk k=1 c Bc2 tk which is a sum of Gaussian random variables hence is Gaussian. Lemma 1.8. Time Inversion Lemma: If B is a Brownian motion then ( Xt := tB 1t 0 t 6= 0 t=0 is also a Brownian motion. Proof. It suffices to check that X has continuous paths, E[Xt ] = 0, E[Xt Xs ] = min{t, s} and that X is Gaussian. N X k=1 λk Xtk = N X k=1 λk tk B t1 k which is the sum of Gaussian random variables hence is Gaussian. E[Xt ] = tE[B 1t ] = 0 since B is a Brownian motion. Let s < t then 1/t < 1/s so E[Xs Xt ] = E[stB 1s B 1s ] = stE[B 1s B 1s ] = st 1 t =s t, B 1t are continuous for t > 0 hence tB 1t has continuous paths for t > 0 so it suffices to show that limt→0+ tB 1t = 0 6 { lim+ Xt = 0} = { lim+ Xq = 0 : q ∈ Q} t→0 by continuity of paths for t q→0 = {∀n ∈ N, ∃m ∈ N : q ∈ Q ∩ (0, 1/m] =⇒ |Xq | < 1/n} \ [ \ = {|Xq | < 1/n} n∈N m∈N q∈Q∩(0,1/m] P( lim+ Xt = 0) = P t→0 \ [ \ {|Xq | < 1/n} n∈N m∈N q∈Q∩(0,1/m] = lim P n→∞ [ \ {|Xq | < 1/n} since the sets are decrea m∈N q∈Q∩(0,1/m] \ = lim lim P {|Xq | < 1/n} n→∞ m→∞ since the sets are increa q∈Q∩(0,1/m] = lim lim lim P (|Xqk | < 1/n) where qk are the ordered elements of Q ∩ (0, 1 n→∞ m→∞ k→∞ = lim lim lim P (|Bqk | < 1/n) n→∞ m→∞ k→∞ = P( lim+ Bt = 0) t→0 Corollary 1.2. If B is a Brownian motion then Bt =0 tα lim t→∞ ∀α > 1/2 Bt lim sup √ = ∞ t t→∞ Bt lim inf √ = −∞ t→∞ t Lemma 1.9. Gaussian Tails: For a > 0 1 a 1 1− 2 a e a a ∞ a 1 − a2 e 2 = a e− ≤ Proof. For z > a > 0 we have that z/a > 1 so Z ∞ Z z2 e− 2 dz ≤ So indeed the upper bound holds. Notice that ∞ Z 2 − a2 Z ∞ e− a 7 z2 2 dz ≤ 1 − a2 e 2 a z − z2 1 a2 e a dz = e− 2 a a x2 2 + 1 − x2 e 2 dx x2 So Z ∞ 2 e − x2 a Z ∞ 1 − a2 1 − x2 2 dx = e − e 2 dx a x2 a Z ∞ 1 a2 x − x2 ≥ e− 2 − e 2 dx a a3 a a2 1 a2 1 = e− 2 − 3 e− 2 dx a a 1 − a2 1 = e 2 1− 2 a a Lemma 1.10. Borel Cantelli: Let {Ai }∞ i=1 ∈ F: P∞ • If i=1 P(Ai ) < ∞ then P(Ai i.o.) = 0 P∞ • If {Ai }∞ i=1 are independent and i=1 P(Ai ) = ∞ then P(Ai i.o.) = 1 Proof. E • "∞ X # χAi = i=1 = ∞ X i=1 ∞ X since χ ≥ 0 E[χAi ] P(Ai ) i=1 <∞ So P a.s. only finitely many of the events occur. P∞ • Let Z = i=1 χAi be the total number of events, then it suffices to show that E[e−Z ] = 0. "∞ # h P∞ i Y − i=1 χAi −χAi E e =E e i=1 = = = ≤ Y ∞E[e−χAi ] i=1 ∞ Y by independence P(Aci ) + e−1 P(Ai ) i=1 ∞ Y (1 − (1 − e−1 )P(Ai )) i=1 ∞ Y e−(1−e −1 )P(Ai ) i=1 = e−(1−e −1 ) P∞ i=1 P(Ai ) =0 Lemma 1.11. Reflection Principle: For a ≥ 0 we have that P(sups≤t Bs > a) = 2P(Bt ≥ a) 8 Proof. Let Ω0 := {sups≤t Bs ≥ a} be the event that the Brownian motion exceeds a before time t. The sets {Bt > a}, {Bt = a}, {Bt < a} form a partition so P(Ω0 ) = P(Ω0 ∩ {Bt > a}) + P(Ω0 ∩ {Bt = a}) + P(Ω0 ∩ {Bt < a}) P(Ω0 ∩ {Bt = a}) = 0 so we don’t need to worry about this event. P({Bt < a}|Ω0 ) = P({Bt > a}|Ω0 ) since one the path reaches the level a there are the same number of paths ending above a as there are below a. This gives us that P(Ω0 ∩ {Bt < a}) = P(Ω0 ∩ {Bt > a}) So indeed we have that P(sups≤t Bs > a) = 2P(Bt ≥ a) Theorem 1.2. p Law of the Iterated Logarithm Let ψ(t) = 2t log(log(t)) then lim sup t→∞ lim inf t→∞ Bt =1 ψ(t) Bt = −1 ψ(t) Proof. By symmetry of Bt it suffices to show just the first limit: lim sup t→∞ To do this we will show that lim supt→∞ • lim supt→∞ Bt ψ(t) Bt ψ(t) Bt =1 ψ(t) ≤ 1 and then lim supt→∞ Bt ψ(t) ≤1 ≤1 Notice that this inequality is equivalent to saying that ∀ε > 0 we have that sufficiently large t. Z ∞ P(Bt > (1 + ε)ψ(t)) = (1+ε)ψ(t) ≤ Bt ψ(t) ≤ 1 + ε for z2 e− 2 √ dz 2π 1 since Bt Gaussian 2 p e−(1+ε) (1 + ε) 4π log(log(t)) log(log(t)) using the Gaussian tails lemma Choose a sequence tn = θn for θ > 1. We want to control the probability that the bound is exceeded at these values tn . P(Bθn > (1 + ε)ψ(θn )) ≤ 2 n 1 p e−(1+ε) log(log(θ )) n (1 + ε) 4π log(log(θ )) 2 1 p = e−(1+ε) log(n log(θ)) (1 + ε) 4π log(n log(θ)) 2 ≤ C(θ, ε)n−(1+ε) where C is some constant depending on θ, ε. By Borel-Cantelli we have that P(Bθn > (1 + ε)ψ(θn ) i.o.) = 0 since ∞ X 2 C(θ, ε)n−(1+ε) < ∞ n=1 hence the Brownian motion will almost surely reach a last n such that at θn it exceeds the bound. 9 We need to show that the process will not exceed the bound between θn , θn+1 for sufficiently large n. P(sups≤θn Bs > (1 + ε)ψ(θn )) = 2P(Bθn > (1 + ε)ψ(θn )) ≤ 2C(θ, ε)n−(1+ε) 2 So by Borel-Cantelli there are a.s. only finitely many intervals [θn , θn+1 ) for which the Brownian exceeds the bound. We therefore have that eventually for t ∈ [θn , θn+1 ): ψ(θn+1 ) Bt ≤ (1 + ε) ψ(t) ψ(θn ) s θn+1 log((n + 1) log(θ)) = (1 + ε) θn log(n log(θ)) s √ log((n + 1) log(θ)) = (1 + ε) θ log(n log(θ)) s lim n→∞ log((n + 1) log(θ)) =1 log(n log(θ)) Moreover we can choose ε arbitrarily small and θ arbitrarily close to 1 hence indeed we have that Bt ψ(t) ≤ 1 a.s. • lim supt→∞ Bt ψ(t) ≥1 Notice that this inequality is equivalent to saying that ∀ε > 0 we have that sufficiently large t. Again set tn = θn for θ > 1 then Bt ψ(t) ≥ 1 − ε for 2 P(Bθn > (1 − ε)ψ(θn )) ≥ C(θ, ε)n−(1−ε) using Gaussian tails. √ Write An = {Bθn − Bθn−1 > (1 − ε) 1 − θ−1 ψ(θn )}. Which are independent by construction, also Bθn − Bθn−1 ∼ N (0, θn − θn−1 ). 2 We have that P(An ) ≥ C(θ, ε)n−(1−ε) , moreover since An are independent we have that by Borel-Cantelli An occurs infinitely often almost surely. This gives us that limsupn→∞ p √ Bθn ≥ (1 − ε) 1 − θ−1 − (1 + ε) θ−1 n ψ(θ ) but we can take ε arbitrarily small and θ arbitrarily large so we have that lim supt→∞ Definition 1.6. Exceptional Point: An exceptional point for the Brownian motion B is a point at which B is differentiable. Definition 1.7. Holder-Continuity: f : [0, ∞) → R is α-Holder continuous for α ∈ (0, 1] at t ∈ [0, ∞) if ∃M, δ > 0 s.t. |ft+s − ft | ≤ M |s|α ∀|s| < δ. Lemma 1.12. If X ∼ N (0, N −1 ), C > 0 then 2C 1 P(|X| ≤ CN −1 ) ≤ √ √ 2π N 10 Bt ψ(t) ≥1 Proof. P(|X| ≤ CN −1 ) = P(X ≤ CN −1/2 ) Z CN −1/2 − z2 e 2 √ dz = 2π −CN −1/2 z2 2C 1 ≤ √ √ maxz e− 2 2π N 2C 1 =√ √ 2π N Proposition 1.1. Let ΩM,δ := {∃t ∈ [0, 1] : |Bt+s − Bt | ≤ M |s| ∀|s| < δ} be the event that B is 1-Holder continuous at some point t ∈ [0, 1]. Then P(ΩM,δ ) = 0. k k+1 , N s.t. |Bt+s − Bs | ≤ M |s| ∀|s| ≤ δ. Proof. Suppose ∃t ∈ N 4 We can choose N sufficiently large such that k+i ∈ [t, t + δ]. We then have that: N i=1 M N 2M |B k+2 − Bt | ≤ N N 3M |B k+3 − Bt | ≤ N N 4M |B k+4 − Bt | ≤ N N |B k+2 − B k+1 | ≤ |B k+2 − Bt | + |B k+1 − Bt | |B k+1 − Bt | ≤ N N |B k+3 N |B k+4 N N N N 3M ≤ N − B k+2 | ≤ |B k+3 − Bt | + |B k+2 − Bt | N N N 5M ≤ N − B k+3 | ≤ |B k+4 − Bt | + |B k+3 − Bt | N N N 7M ≤ N Since B is a Brownian motion we have that B k+i+1 − B k+i ∼ N (0, N −1 ) N N for i = 1, 2, 3. 11 So P(ΩM,δ ) ≤ P N[ −3 \ 3 |B k+i+1 N k=1 i=1 (2i − 1)M − B k+i | ≤ N N ! N −3 Y 3 X (2i − 1)M ≤ P |B k+i+1 − B k+i | ≤ N N N i=1 k=1 ≤ N −3 X k=1 C N 3/2 C ≤√ N C lim √ = 0 N →∞ N So indeed since N can be chosen arbitrarily large we have that P(ΩM,δ ) = 0 Proposition 1.2. Let α > 1/2 and ΩM,δ := {∃t ∈ [0, 1] : |Bt+s − Bt | ≤ M |s|α ∀|s| < δ} be the event that B is α-Holder continuous at some point t ∈ [0, 1]. Then P(ΩM,δ ) = 0. k k+1 , N s.t. |Bt+s − Bs | ≤ M |s|α ∀|s| ≤ δ. Proof. Suppose ∃t ∈ N Let ρ = d4/αe. ρ We can choose N sufficiently large such that k+i ∈ [t, t + δ]. We then have that: N i=1 iM Nα − B k+i | ≤ |B k+i+1 − Bt | + |B k+i − Bt | |B k+i − Bt | ≤ N |B k+i+1 N N N N (2i − 1)M ≤ Nα Since B is a Brownian motion we have that B k+i+1 − B k+i ∼ N (0, N −1 ) N N for i = 1, ..., ρ. So P(ΩM,δ ) ≤ P N −ρ+1 [ ρ−1 \ ≤ |B k+i+1 N i=1 k=1 NX −ρ+1 ρ−1 Y P |B k+i+1 N k=1 ≤ i=1 NX −ρ+1 k=1 (2i − 1)M − B k+i | ≤ N Nα ! (2i − 1)M − B k+i | ≤ N Nα C N α(ρ−1)/2 C ≤√ N C lim √ = 0 N →∞ N So indeed since N can be chosen arbitrarily large we have that P(ΩM,δ ) = 0 12 Corollary 1.3. P ∞ [ ∞ [ ! ΩM, N1 =0 M =1 N =1 Theorem 1.3. Kolmogorov Continuity Theorem: Suppose {Xt }t∈[0,1] has continuous paths and satisfies E[|Xt − Xs |p ] ≤ C|t − s|1+γ for some γ, p > 0 then X has α-Holder continuous paths for α ∈ (0, γ/p). Corollary 1.4. A Brownian motion has α-Holder continuous paths for α < 1/2. Proof. Let B be a Brownian motion then E[|Bt − Bs |p ] = E[|X|p ] X ∼ N (0, t − s) p p/2 = E[|Y | ]|t − s| Y ∼ N (0, 1) p/2 = Cp |t − s| So the Kolmogorov continuity theorem holds for a Brownian motion with α ≤ γ/p where γ + 1 = p/2 hence the theorem holds for α < 1/2. Lemma 1.13. Markov’s Inequality: If Z ≥ 0 is a random variable then P(Z ≥ 0) ≤ E[Z] a Proof. E[Z] ≥ E[ZIZ≥a ] since Z ≥ 0 ≥ E[aIZ≥a ] = aE[IZ≥a ] = aP(Z ≥ a) Corollary 1.5. Suppose {Xt }t∈[0,1] has continuous paths and satisfies E[|Xt − Xs |p ] ≤ C|t − s|1+γ for some γ, p > 0 then P(|Xt − Xs | ≥ a) ≤ C|t − s|1+γ ap Proof. P(|Xt − Xs | ≥ a) = P(|Xt − Xs |p ≥ ap ) ≤ E[|Xt − Xs |p ] ap C|t − s|1+γ ≤ ap by Markov by KCT 13 Proposition 1.3. For α > 1/2 we have that P(B α − Holder Continuous ∀t ≥ 0) = 1 Proof. Let AN = 2N [ |X k 2n − X k−1 |> n 2 k=1 1 2nα be the event that some interval has a jump larger than 2−N α then 2N X P(AN ) ≤ P |X k 2n 1 − X k−1 |> n 2nα 2 k=1 ≤ 2N C k − X 2N 1+γ k−1 2N 2−N αp k=1 ≤ 2N C 2kN − 1+γ k−1 2N 2−N αp −N (γ−αp) = C2 Since B is a Brownian motion and we have assumed α < 1/2 this indeed holds by corollary 1.4. By Borel Cantelli we have that ∞ X P(AN ) < ∞ N =1 hence AN occurs only finitely often and hence we can find N sufficiently large such that B is α-Holder continuous between fixed points of distance 2−N . So we have that on some countable grid of points D = {x ∈ Q : x = k/2m , k, m ∈ N} we have that the Brownian motion is α-Holder continuous. We need to extend this to the entire space but this follows by continuity of paths. 2 Markov Processes Definition 2.1. Independent: • The σ-Fields F, G on (Ω, P) are independent if P(A ∩ B) = P(A)P(B) ∀A ∈ F, B ∈ G • Random variables X, Y are independent if P(X ∈ A, Y ∈ B) = P(X ∈ A)P(Y ∈ B) Lemma 2.1. Random variables X, Y are independent iff E[f (X)g(Y )] = E[f (X)]E[g(Y )] ∀f, g bounded measurable functions. Proof. Suppose E[f (X)g(Y )] = E[f (X)]E[g(Y )] ∀f, g bounded measurable functions. 14 ∀A, B measurable Then let A ∈ F, B ∈ G and set f = IA , g = IB which are bounded and measurable. Moreover P(X ∈ A, Y ∈ B) = E[IA (X)IB (Y )] = E[IA (X)]E[IB (Y )] = P(X ∈ A)P(Y ∈ B) Now let f, g be bounded measurable functions. There exist sequence of simple functions fn = n X λi IAi gm = i=0 m X γj IBj j=0 s.t. limn→∞ fn = f and limn→∞ gn = g so by mechanics of measure theory the result follows: E[f (X)g(X)] = E[ lim fn (X) lim gm (Y )] n→∞ m→∞ = lim lim E[fn (X)gm (Y )] n→∞ m→∞ = lim lim n→∞ m→∞ = lim lim n→∞ m→∞ = lim lim n→∞ m→∞ = lim lim n→∞ m→∞ n X m X i=0 j=0 n X m X i=0 j=0 n X m X i=0 j=0 n X m X λi γj E[IAi (X)IBj (Y )] λi γj P(X ∈ Ai , Y ∈ Bj ) λi γj P(X ∈ Ai )P(Y ∈ Bj ) λi γj E[IAi (X)]E[IBj (Y )] i=0 j=0 " = lim lim E n→∞ m→∞ n X m X λi IAi (X) E γj IBj (Y ) # i=0 j=0 = E[ lim fn (X)]E[ lim gm (Y )] n→∞ m→∞ = E[f (X)]E[g(Y )] Definition 2.2. σ-Field Generated by X: For a random variable X : Ω → R we say that σ(X) := {X −1 (A) : A ∈ B(R)} is the σ-field generated by X. Corollary 2.1. For random variables X, Y we have that X ` Y ⇐⇒ σ(X) ` σ(Y ) Theorem 2.1. Markov Property for Brownian Motion: If B is a Brownian motion and t0 > 0 define Xt := Bt0 +t − Bt0 . Then we have that a σ({Xt }t≥0 ) σ({Bt }t∈[0,t0 ) ) Definition 2.3. π-System: A collection A of subsets is called a π-system if it is closed under finite intersections. Lemma 2.2. If F0 , G0 are π-systems such that P(A ∩ B) = P(A)P(B) ∀A ∈ F0 , B ∈ G0 then P(A ∩ B) = P(A)P(B) ∀A ∈ σ(F0 ), B ∈ σ(G0 ) 15 Corollary 2.2. A Brownian motion has the Markov property. ` Proof. It suffices to check that {Xtk }N {Bsr }M r=1 so k=1 h PM i h PM i PN PN E e r=1 λr Bsr e k=1 µk Xtk = E e r=1 λr Bsr e k=1 µk BT +tk −BT ) h PM i PN = E e r=1 λ̂r (Bsr −Bsr−1 )+ k=1 µ̂k (BT +tk −BT +tk−1 ) h PM i h PN i = E e r=1 λ̂r (Bsr −Bsr−1 ) E e k=1 µ̂k (BT +tk −BT +tk−1 ) By independent increments. Definition 2.4. Local Maxima: t ∈ R+ is the local maxima of a Brownian motion B if ∃ε > 0 s.t. Bt > Bs ∀s ∈ (t − ε, t + ε) \ {t} Proposition 2.1. If B is a Brownian motion then: • a < b < c < d implies that P(max[a,b] Bt 6= max[c,d] Bt ) = 1 • P(∃1 t∗ ∈ [a, b] : Bt∗ = max[a,b] Bt ) = 1 • Local maxima are dense in [0, ∞) Proof. • Write Xt = Bc+t − Bc Define Y = max[c,d] Bt − Bc = max[0,d−c] Xt which is independent of σ(Bs : s ≤ c) by the Markov property. Similarly define Z = max[a,b] Bt − Bc Then it suffices to show that P(Y 6= Z) = 1 By the Markov property we have that Y, Z are independent, moreover Y has a density so let ν(dy) = P(Y ∈ dy), µ(dz) = P(Z ∈ dz) then for ϕ : R2 → R we have Z Z E[ϕ(Y, Z)] = ϕ(y, z)ν(dy)µ(dz) By independence so let ϕ(y, z) = Iy=z P(Y = Z) = E[Iy=z ] Z Z = Iy=z ν(dy)µ(dz) =0 • By contradiction suppose ∃t1 6= t2 ∈ [a, b] such that Bt1 = Bt2 = max[a,b] Bt Then we can find c, d ∈ Q such that a < t1 < c < d < t2 < b. But P(max[a,c] Bt = max[d,b] Bt ) = 0 for any fixed a, b, c, d and there are only coutably many such events satisfying the above condition we must have that [ P( max[a,c] Bt = max[d,b] Bt ) = 0 c∈Q∩[a,b] • Choose interval [a, a + n−1 ] for some n ∈ N, a ∈ Q By the previous part P(∃1 t∗ ∈ [a, a + n−1 ] : Bt∗ = max[a,a+n−1 ] Bt ) = 1 Any interval contains some such interval of the form [a, a + n−1 ] and since there are countably many we indeed have that local maxima are dense. 16 Definition 2.5. Brownian Filtration: For a Brownian motion B the Brownian filtration at t is defined as FtB := σ(Bs : s ≤ t) Definition 2.6. Germ Field: For a Brownian motion B the germ field at t is defined as \ FtB+ := FsB s>t Corollary 2.3. An alternative version of the Markov property is that for fixed T ∈ R we have that when Xt := BT +t = BT we have a σ(Xt : t ≥ 0) FTB+ Corollary 2.4. F0B+ is P-trivial. ` Proof. For T = 0 we have that {Bt : t ≥ 0} F0B+ ` So F0B+ F0B+ which can only occur if F0B+ is P-trivial. ` Theorem 2.2. σ({Xs }s≥0 ) FtB+ where Xs := Bt+s − Bt Proof. Take A ∈ FtB+ and B ∈ σ({Xs }s≥0 ) It suffices to choose events in π-systems generating these σ-fields so choose B = {(Xs1 ∈ C1 , ..., Xsm ∈ Cm )} Now it suffices to check that h Pm i h Pm i E ei k=1 λk Xsk eiθχA = E ei k=1 λk Xsk E eiθχA For ε > 0 let Xsε = Bt+s+ε − Bt+ε then Xsε ` FtB+ for any ε > 0. h Pm i h Pm i ε ε E ei k=1 λk Xsk eiθχA = E ei k=1 λk Xsk E eiθχA h Pm i h Pm i ε lim E ei k=1 λk Xsk E eiθχA = E ei k=1 λk Xsk E eiθχA ε→0 h Pm i h Pm i ε lim E ei k=1 λk Xsk eiθχA = E ei k=1 λk Xsk eiθχA ε→0 By dominated convergence theorem and continuous paths so indeed the statement holds. 3 Markov Transition Functions Definition 3.1. Markov Transition Kernal: p : (0, ∞) × Rd × B(R) → [0, 1] is a Markov transition kernal if • A → pt (x, A) is a probability. • x → pt (x, A) is measurable. Definition 3.2. Markov Process: {Xt }t≥0 process on Rd with transition kernal pt started from x is a Markov process if ! Z Z Y n n \ P {Xti ∈ Ai } = ... pti −ti−1 (xi−1 , dxi ) i=1 A1 An i=1 where x0 := x, t0 := 0. 17 Corollary 3.1. From simple measure theory we have that for f : Rn → R and Markov process X we get Z Z n Y E[f (Xt1 , ..., Xtn )] = ... f (x1 , ..., xn ) pti −ti−1 (xi−1 , dxi ) A1 An i=1 Btx Lemma 3.1. := x + Bt is a Markov process started at x with kernal pt (x, dy) = qt (y, x)dy where qt (z) is the Gaussian density with mean zero and variance z. Proof. P n \ ! {Btxi ∈ dxi } n \ =P i=1 ! {Btxi − Btxi−1 ∈ d(xi − xi−1 )} i=1 = = n Y i=1 n Y P(Btxi − Btxi−1 ∈ d(xi − xi−1 )) by independence qti −ti−1 (xi − xi−1 ) i=1 By integrating both sides over Ai we have the required result. Lemma 3.2. Process {Xt }t≥0 with kernal p is a Markov process iff ∀s < t we have P(Xt ∈ A|FsX ) = pt−s (xs , A) = P(Xt−s ∈ A|X0 = xs ) Remark 3.1. For a probability space (Ω, F, P) we have that L2 (Ω, F, P) := {X : Ω →p R|E[X 2 ] < ∞} is an inner product space with inner product < X, Y >= E[XY ] and norm ||X||L2 := E[X 2 ] Definition 3.3. Conditional Expectation: For probability space (Ω, F, P) and sub-σ-field G ⊆ F we say that Y = E[X|G] if < X − Y, Z >= 0 whenever Z ∈ L2 (Ω, G, P) Intuitively E[X|G] is the closest G measurable function to X. Lemma 3.3. For X ∈ L2 (Ω, F, P) ∃1 Y ∈ L2 (Ω, G, P) s.t. < X − Y, Z >= 0 ∀Z ∈ L2 (Ω, G, P) Corollary 3.2. For X ∈ L2 (Ω, F, P) and G ⊆ F sub σ-field TFAE • < X − Y, Z >= 0 ∀Z ∈ L2 (Ω, G, P) • E[(X − Y )Z] = 0 ∀Z ∈ L2 (Ω, G, P) • E[XZ] = E[Y Z] ∀Z ∈ L2 (Ω, G, P) • E[XχA ] = E[Y χA ] ∀A ∈ G R R • A X(ω)dP(ω) = A Y (ω)dP(ω) ∀A ∈ G ` Lemma 3.4. Suppose X is G measurable, Z G and ϕ : R2 → R is measurable then: E[ϕ(X, Z)|G] = E[ϕ(x, Z)]|x=X Proposition 3.1. If (Ω, F, P) is a probability space and G, H are sub σ-algebras of F then: • E[E[X|G]] = E[X] • If X is G-measurable then E[X|G] = X a.s. • E[a1 X1 + a2 X2 |G] = a1 E[X1 |G] + a2 E[X2 |G] • If X ≥ 0 then E[X|G] ≥ 0 a.s. 18 • If Xn ≥ 0 is an increasing sequence of random variables converging to X a.s. then E[Xn |G] converges to E[X|G] a.s. • If Xn is a sequence of random variables then E[lim inf n→∞ Xn G] ≤ lim inf n→∞ E[Xn |G] a.s. • If |Xn (ω)| ≤ |V (ω)| ∀n such that E[V ] < ∞ and Xn converges to X a.s. then E[Xn |G] converges to E[X|G] • If c : R → R is convex and E[|c(X)|] < ∞ then E[c(X)|G] ≥ c(E[X|G]) • If H ⊆ G then E[E[X|G]|H] = E[X|H] a.s. • If Z is G measurable and bounded then E[ZX|G] = ZE[X|G] ` • If H σ(σ(X), G) then E[X|σ(G, H)] = E[X|G] a.s. Lemma 3.5. X is a Markov process iff Z E[f (Xt1 , ..., Xtn )] = Z ... f (xt1 , ..., xtn ) n Y pti −ti−1 (xti−1 , dxti ) i=1 For any f measurable. Proof. By measure theory it suffices to show this for f (Xt1 , ..., Xtn ) = n Y ϕi (Xti ) i=1 for {ϕi }ni=1 measurable. By induction: ## " n # " " n Y Y X E ϕi (Xti ) = E E ϕi (Xti )Ftn−1 i=1 i=1 =E "n−1 Y # ϕi (Xti )E[ϕn (Xtn )|FtXn−1 ] i=1 =E "n−1 Y # Z ϕi (Xti ) ϕi (xtn )ptn −tn−1 (xtn−1 , dxtn ) i=1 Z = ... Z Y n ϕi (xti )pti −ti−1 (xti−1 , dxti ) i=1 4 Strong Markov Property Definition 4.1. Filtration: Collection of σ-fields {Ft }t≥0 is a filtration on (Ω, F, P) if Fs ⊆ Ft ⊆ F for any s ≤ t Definition 4.2. Stopping Time: T : Ω → R+ is a stopping time for filtration {Ft }t≥0 if {T ≤ t} ∈ Ft Definition 4.3. Hitting Time: For A ⊆ R we write the hitting time of the Brownian motion B to be TA := inf {t ≥ 0 : Bt = a} 19 ∀t ∈ R+ Lemma 4.1. If C is closed then {TC ≤ t} ∈ FtB Proof. {TC ≤ t} = ∞ \ [ {infr∈C ||Bq − r|| < 1/n} ∈ FtB n=1 q∈Q∩[0,t] Lemma 4.2. For stopping time T with respect to filtration {Ft }t≥0 then for FT := {A : A ∩ {T ≤ t} ∈ Ft ∀t ≥ 0} we have that • FTB is a σ-field • If S ≤ T are stopping times with respect to {Ft }t≥0 then FS ⊆ FT • T is FT measurable. Proof. • φ ∩ {T ≤ t} = φ ∈ Ft Ω ∩ {T ≤ t} = {T ≤ t} ∈ Ft A ∈ FT =⇒ Ac ∩ {T ≤ t} = {T ≤ t} \ A ∈ Ft T∞ T∞ {Ai }∞ ∈ F =⇒ ( A ) ∩ {T ≤ t} = t i=1 i=1 i i=1 (Ai ∩ {T ≤ t}) ∈ Ft • A ∩ {T ≤ t} = A ∩ {T ≤ t} ∩ {S ≤ t} So if A ∩ {S ≤ t} ∈ Ft then by definition {T ≤ t} ∈ Ft which is a σ-field so A ∩ {T ≤ t} ∩ {S ≤ t} ∈ Ft • It suffices to check that {T ≤ s} ∈ FT ∀s ∈ R by the π-systems lemma. {T ≤ s} ∩ {T ≤ t} = {T ≤ min{s, t}} ∈ Fmin{s,t} ⊆ Ft Theorem 4.1. Strong Markov Property: If B is a Brownian motion, FtB = σ({Bs }s≤t ) and T is a stopping time with respect to {FtB }t≥0 where T < ∞ a.s. then: X := BT +s − BT is a Brownian motion independent of FTB . Proof. We will split the proof into two cases: • Suppose T takes only a discrete set of values ` Ω= We want to show that Xt : +BT +t − BT FTB Choose A ∈ FTB and let F be measurable. E[F ({BT +t − BT }t≥0 )χA ] = ∞ X S∞ k=1 {T = tk } E[χA∩{T =tk } F ({Btk +t − Btk }t≥0 )] k=1 But, buy the Markov property we have that F ({Btk +t − Btk }t≥0 ) ` Ftk and A ∩ {T = tk } = (A ∩ {T ≤ tk }) \ (A ∩ {T ≤ tk−1 }) ∈ FtBk So χA∩{T ≤tk } ` F ({Btk +t − Btk }t≥0 ) hence we have that E[F ({BT +t − BT }t≥0 )χA ] = = = ∞ X k=1 ∞ X k=1 ∞ X E[χA∩{T =tk } F ({Btk +t − Btk }t≥0 )] E[χA∩{T =tk } ]E[F ({Btk +t − Btk }t≥0 )] E[χA∩{T =tk } ]E[F ({Bt }t≥0 )] k=1 = E[χA ]E[F ({Bt }t≥0 )] So indeed the proof holds when T can only take a countable number of values. 20 • For general T we want to approximate T using a discrete set of stopping times so define TN := ∞ X j IT ∈[ j−1 , j ) N N N j=1 which is the smallest multiple of 1/N strictly larger than T . Notice that limN →∞ TN = T and that TN > T for any N . j We need to check that TN is a stopping time so let t ∈ j−1 N , N {TN j−1 ≤ t} = TN ≤ N j−1 = T < N [ B {T ≤ s} ∈ F B = j−1 ⊆ Ft N j−1 s∈Q∩[0, N ) So indeed TN is a stopping time hence from the previous case we have that {BTN +t − BTN }t≥0 is a Brownian motion independent of FTBN . Moreover TN > T so FTB ⊆ FTBN . Let A ∈ FTB then h PM i h PM i lim E ei k=1 θk (BTN +sk −BTN ) P(A) = lim E ei k=1 θk (BTN +sk −BTN ) χA N →∞ N →∞ h PM i = E ei k=1 θk (BT +sk −BT ) χA h PM i = E ei k=1 θk (BT +sk −BT ) P(A) So indeed we have that BT +s − BT is a Brownian motion independent of FTB . Proposition 4.1. If c ≤ b ≤ a then Ta − Tb ` Tc and Ta − Tb has the same distribution as Ta−b Proof. Xt := BTb +t − b is a Brownian motion independent of FTBb by the strong Markov property. Since Tc is measurable with respect to FTc ⊆ FTb we have that Tc is independent of X which is a measurable function of {Bt }t≥Tb . Moreover Ta − Tb = inf {t : Xt = a − b} which clearly has the same distribution as Ta−b Definition 4.4. Isolated Point: For a set A we say that t ∈ A is isolated if ∃ε > 0 such that (t − ε, t + ε) ∩ A = {t} Lemma 4.3. If A ⊆ R \ φ is closed and has no isolated points then it is uncountable. Proof. {0, 1}N is uncountable, in particular we can choose the set of sequences with entries 0, 1 which do not end with an infinite sequence of zeros. By doing this every element is unique and the set is still countable. We want to form a bijection between this set and A. Since A 6= φ and has no isolated points we can choose t0 6= t1 ∈ A. Write B0 = B(t0 , ε0 ) ∩ A, B1 = B(t1 , ε1 ) ∩ A where ε1 are chosen small enough that B0 ∩ B1 = φ. Since A has no isolated points neither do B0 , B1 so we can find t00 , t01 ∈ B0 , t10 , t11 ∈ B1 . We can then choose ε00 , ε01 , ε10 , ε11 small enough such that B00 = B(t00 , ε00 ) ∩ A, B01 = B(t01 , ε01 ) ∩ A, B10 = B(t10 , ε10 ) ∩ A, B11 = B(t11 , ε11 ) ∩ A are disjoint and have no isolated points. 21 Repeating this inductively and choosing values of ε → 0 we have that a = (a0 , a1 , ...) can be mapped to a unique point ∞ \ ta = Ba0 ,...,ai i=0 Proposition 4.2. Let Z := {T : Bt = 0} then • L(Z) = 0 a.s. • @ an isolated point a.s. • Z is uncountable a.s. Proof. • Z ∞ L(Z) = χt∈Z dt Z0 ∞ χBt =0 dt = 0 Z E[L(Z)] = ∞ P(Bt = 0)dt by Tonelli 0 =0 So indeed L(Z) = 0 a.s. • Let τs := inf {t ≥ s : Bt = 0} and Ẑ = {τs : s ∈ Q} Bτs +t is a Brownian motion by the strong Markov property so by the law of the iterated logarithm cannot be isolated. Suppose τ ∈ Z \ Ẑ then ∃{Sn }∞ n=1 ∈ Q converging to τ from the left each with corresponding τSn < τ . But Sn ≤ τSn < τ so we must have that τsn converges to τ and hence τ cannot be isolated. • If A is closed and has no isolated points then it is uncountable. Z := {t ≥ 0 : Bt = 0} is closed by continuity of B and hence Z is uncountable. Lemma 4.4. Arcsine Laws For Brownian Motion: If B is a Brownian motion for the interval [0, 1] then • Let M be the unique time where B attains its maximum BM = supt∈[0,1] Bt then P(M ∈ dt) = 1 π p t(1 − t) • Let T be the final time where B attains the value zero T = sup{t ∈ [0, 1]|Bt = 0} then P(T ∈ dt) = • Let L be the time B spends above zero L = R1 0 1 π p t(1 − t) χBs >0 ds then P(L ∈ dt) = π 1 p t(1 − t) Theorem 4.2. Levy’s Theorem: If B is a Brownian motion and St := sups≤t Bs then Xt := St − Bt defines a reflected Brownian motion. 22 5 Brownian Martingales Definition 5.1. Martingale: For a filtration {Ft }t≥0 on probability space (Ω, F, P) a process {Mt }t≥0 is called a martingale if: • E[|Mt |] < ∞ • {Mt }t≥0 is adapted to {Ft }t≥0 • ∀s ≤ t E[Mt |Fs ] = Ms Theorem 5.1. Optional Stopping Theorem (version 1): If {Mt }t≥0 is a bounded martingale on {Ft }t≥0 and T a bounded stopping time then E[MT ] = E[M0 ] Proof. We shall split the proof into two cases: • Case 1) ∃{ti }ni=1 ∈ R+ such that ti < ti+1 ≤ K ∀i and T ∈ {ti }ni=1 . E[MT ] = = n X k=1 n X E[MT χT =tk ] E[Mk χT =tk ] k=1 = E[Mk ] = E[M0 ] • Case 2) For general T . Define Tn = N X j χT ∈[ j−1 , j ) N N N j=1 which is a finite sequence of bounded stopping times and therefore satisfy case 1. TN is a decreasing sequence in N such that limN →∞ TN = T so limN →∞ MTN = MT by continuity of paths. By DCT we have that: E[MT ] = lim E[MTN ] N →∞ = E[M0 ] Corollary 5.1. For a < 0 < b we have that P(Ta < Tb ) = b b−a where Ta = inf {t ≥ 0 : Bt = a}, Tb = inf {t ≥ 0 : Bt = b} Proof. Bt is a martingale and T := min{Ta , Tb } is a stopping time so for fixed N we have that T ∧ N is a bounded stopping time. By the optional stopping theorem we have that E[BT ∧N ] = 0 23 |BT ∧N | ≤ max{|b|, |a|} for any N so by dominated convergence theorem we have that E[BT ] = lim E[BT ∧N ] = 0 N →∞ Which gives that: 0 = E[BT ] = aP(Ta < Tb ) + bP(Tb < Ta ) = aP(Ta < Tb ) + b(1 − P(Ta < Tb )) = b + P(Ta < Tb )(a − b) b P(Ta < Tb ) = b−a Corollary 5.2. If T = min{Ta , Tb } for a < 0 < b then E[T ] = |ab| Proof. E[Bt2 − t|Fs ] = E[(Bt − Bs )2 − Bs2 + 2Bs Bt − t|Fs ] = E[(Bt − Bs )2 + 2Bs (Bt − Bs ) + Bs2 |Fs ] = t − s + Bs2 − t = Bs2 − s hence Bt2 − t is a martingale so by the optional stopping theorem E[BT2 ∧N − T ∧ N ] = 0 |BT2 | ≤ max{b2 , a2 } so by dominated convergence theorem limN →∞ E[BT2 ∧N ] = E[BT2 ] Furthermore by monotone convergence theorem we have that limN →∞ E[T ∧ N ] = E[T ] So we get that E[BT2 ] − E[T ] = lim E[BT2 ∧N ] − E[T ∧ N ] N →∞ = lim E[BT2 ∧N − T ∧ N ] N →∞ =0 E[T ] = E[BT2 ] = b2 P(Tb < Ta ) + a2 P(Ta < Tb ) a2 b b2 a + b−a b−a = −ab =− Corollary 5.3. If θ > 0 then E[e− θ2 2 Ta 24 ] = eθa Proof. eθBt − θ2 2 t is a martingale and Ta ∧ N is a bounded stopping time hence E[eθBTa ∧N − Notice that e− that θ2 2 Ta ∧N θ2 2 Ta ∧N ] = E[eθB0 ] = 1 ≤ 1 for any N and eθBTa ∧N ≤ eθa so by dominated convergene theorem we have E[eθBTa − θ2 2 Ta ] = lim E[eθBTa ∧N − θ2 2 Ta ∧N N →∞ ]=1 and since BTa = a we indeed have that E[e− θ2 2 Ta ] = eθa Corollary 5.4. Let Xt := Bt − ct and Ta := inf {t ≥ 0 : Xt = a} then P(Ta < ∞) = e−2ca Proof. For θ > 0 we have that eθBt − eθBt − θ2 2 t = eθ(Xt +ct)− =e θ2 2 θ2 2 t is a martingale. Moreover t 2 θXt + c− θ2 t is a martingale on FtX . Ta is a stopping time so Ta ∧ N is a bounded stopping time. eθXTa ∧N ≤ eθa since θ > 0 moreover lim eθXTa ∧N = eθa χTa <∞ N →∞ e 2 θc− θ2 Ta ∧N θ2 2 ≤ 1 so long as θc < moreover lim e 2 θc− θ2 Ta ∧N N →∞ =e 2 θc− θ2 Ta χTa <∞ So by dominated convergence theorem we have that E[eθa χTa <∞ e 2 θc− θ2 Ta ] = lim E[e 2 θXTa ∧N + θc− θ2 Ta ∧N N →∞ ] = E[eθX0 ] = 1 So by choosing θ = 2c we have that P(Ta < ∞) = E[χTa <∞ ] θa = E[e χTa <∞ e 2 θc− θ2 Ta ]e−θa = e−θa = e−2ca Proposition 5.1. Lots of Brownian Martingales: ∂f ∂2f If B is a d-dimensional Brownian motion, f ∈ C2,1 and f, ∂f ∂t , ∂xi , ∂xi ∂xj are of at most exponential growth then Z t ∂f x 1 f (Btx , t) − (Bs , s) + ∆f (Bsx , s)ds 2 0 ∂s is a martingale. 25 Proof. • f has at most exponential growth hence ∃C0 , C1 s.t. x f (Btx , t) ≤ C0 eC1 |Bt | So we have that x E[|f (Btx , t)|] ≤ C0 E[eC1 |Bt | ] x x < C0 E[eC1 Bt + e−C1 Bt ] C 2 t2 C 2 t2 −C1 x+ 12 C1 x+ 12 +e = C0 e <∞ Similarly ∂f ∂t , ∆f are of at most exponential growth so ∃C2 , C3 s.t. x ∂f x 1 (B , s) + ∆f (Bsx , s) ≤ C3 eC4 Bs ∂s s 2 In particular we have that Z t Z t x 1 ∂f x (Bs , s) + ∆f (Bsx , s)ds ≤ E C3 eC4 Bs ds E 2 0 0 ∂s Z t 2 s2 2 s2 C4 −C4 ≤ C3 eC4 x+ 2 + eC4 x+ 2 ds 0 ≤ tC3 e2C4 |x|+ 2 t2 C4 2 <∞ So indeed Z E[|Mt |] ≤ E[|f (Btx , t)|] + E t 0 ∂f x 1 (Bs , s) + ∆f (Bsx , s)ds < ∞ ∂s 2 • WLOG let x = 0 then the original case is a shift of this one and since we have at most exponential growth this will merely be a change in parameters. Moreover let Lf := ∂f 1 + ∆f ∂t 2 then E[Mt |FsB ] Z t B = E f (Bt , t) − Lf (Br , r)dr|Fs 0 Z s Z t B = E f (Bs , s) + f (Bt , t) − f (Bs , s) − Lf (Br , r)dr − Lf (Br , r)dr|Fs 0 s Z t−s = Ms + E f (Bs + Xt−s , t) − f (Bs , s) − Lf (Bs + Xs+r , s + r)dr|FsB Xr = Bs+r − Bs 0 Z t−s = Ms + E f (z + Xt−s , t) − f (z, s) − Lf (Xr + z, s + r)dr 0 z=Bs The result follows if we can show that the arguement of the expectation is null, in particular defining g(y, t) := f (z + y, t + s) we want to show that Z t E[g(Xt , t)] − g(0, 0) = E Lg(Xr , r)dr 0 26 By Fubini since E[|Mt |] < ∞ this is equivalent to showing Z t E[g(Xt , t)] − g(0, 0) = E [Lg(Xr , r)] dr 0 Differentiating both sides gives us that this is equivalent to d E[g(Xt , t)] = E[Lg(Xt , t)] dt So letting |x|2 e− 2t φ(x, t) = (2π)d/2 gives Z d d E[g(Xt , t)] = g(x, t)φ(x, t)dx dt dt Rd Z ∂φ ∂g = +g (x, t)dx φ ∂t ∂t Rd Z ∂g 1 = φ + g ∆φ (x, t)dx ∂t 2 Rd Z ∂g 1 = φ(x, t) + ∆g (x, t)dx ∂t 2 Rd ∂φ 1 = ∆φ ∂t 2 Using Integration by Parts = E[Lg(Xt , t)] If f is such that ∆f = 0 or ∂f ∂s + 12 ∆f = 0 then f (Btx , t) is a martingale. Theorem 5.2. Dirichlet Problem: Suppose D ⊂ Rd is open and bounded and u ∈ C 2 (Rd ) solves ( ∆u(x) = 0 x∈D u(x) = f (x) x ∈ ∂D for some boundary temperature f then u(x) = E[f (BTx )] where T = inf {t ≥ 0 : Btx ∈ ∂D}. Proof. By lots of Brownian motion we have that Z x u(Bt ) − t 0 1 ∆u(Bsx )ds 2 is a martingale since we can modify u outside D to ensure the derivatives are of at most exponential growth. By optional stopping theorem we have that " # Z T ∧N 1 x x ∆u(Bs )ds u(x) = E u(BT ∧N ) − 2 0 By the first condition on u we have that Z T ∧N 0 1 ∆u(Bsx )ds = 0 2 Moreover u is bounded on D since D is bounded and u is twice differentiable. By DCT we have that u(x) = lim E[u(BTx ∧N )] = E[f (BTx )] N →∞ 27 Corollary 5.5. If B is a d dimensional Brownian motion for d ≥ 3 then P(Ta < Tb ) = 1 1 − bd−2 |x|d−2 1 1 − bd−2 ad−2 Where Ta = inf {t ≥ 0 : |Btx | = a}, Tb = inf {t ≥ 0 : |Btx | = b} for 0 < a < |x| < b. Proof. Let u(x) = |x|−(d−2) then ∆u(x) = 0 whenever x ∈ Rd \ {0}. Write D = {x ∈ Rd : a < |x| < b} and consider the Dirichlet problem for T = Ta ∧ Tb . This gives us that: 1 u(x) = E |Btx |d−2 1 1 = d−2 P(Ta < Tb ) + d−2 P(Tb < Ta ) a b 1 1 = d−2 P(Ta < Tb ) + d−2 (1 − P(Ta < Tb )) a b By the definition of u(x) we therefore have by rearranging P(Ta < Tb ) = 1 1 − bd−2 |x|d−2 1 1 − bd−2 ad−2 Corollary 5.6. For d ≥ 3 we have that P(Ta < ∞) = a |x| d−2 Corollary 5.7. For d ≥ 3 we have that |Bt | → ∞. Proof. If |Bt | 9 ∞ then ∃K < ∞ and {tN }∞ N =1 such that |BtN | ≤ K. Let TN = inf {t ≥ 0 : |Bt | ≥ N } then by the law of the iterated logarithm we have that TN < ∞ a.s. (N ) If we define Xt := BTN +t − BTN which is a Brownian motion be strong Markov we have that (N ) P(Xt ∈ B(0, K) eventually) = K N d−2 from the previous corollary so P \ (N ) {Xt ∈ B(0, K) eventually} = 0 N ≥K In particular P ∞ [ \ (N ) {Xt ∈ B(0, K) eventually} = 0 K=1 N ≥K Corollary 5.8. If B is a 2 dimensional Brownian motion then P(Ta < Tb ) = log(b) − log(|x|) log(b) − log(a) Where Ta = inf {t ≥ 0 : |Btx | = a}, Tb = inf {t ≥ 0 : |Btx | = b} for 0 < a < |x| < b. 28 Proof. Let u(x) = log(|x|) then ∆u(x) = 0 whenever x ∈ Rd \ {0}. Write D = {x ∈ Rd : a < |x| < b} and consider the Dirichlet problem for T = Ta ∧ Tb . This gives us that: u(x) = E[log(Btx )] = log(a)P(Ta < Tb ) + log(b)P(Tb < Ta ) = log(a)P(Ta < Tb ) + log(b)(1 − P(Ta < Tb )) By the definition of u(x) we therefore have by rearranging P(Ta < Tb ) = log(b) − log(|x|) log(b) − log(a) Corollary 5.9. A 2 dimensional Brownian motion hits and non-empty open ball with probability 1 but the probability it hits a specific point is 0. Furthermore the 2 dimensional Lebesgue measure of {Bt }t≥0 is 0 but B is dense in R2 . Theorem 5.3. Poisson Problem: Suppose D ⊂ Rd is bounded and u ∈ C 2 (Rd ) solves ( 1 2 ∆u(x) = −g(x) u(x) = 0 x∈D x ∈ ∂D for some interior temperature function g then "Z # T g(Bsx )dx u(x) = E 0 where T = inf {t ≥ 0 : Btx ∈ ∂D}. Proof. By modifying u outside D we can satisfy the lots of Brownian martingales proposition with Z t 1 x u(Bt ) − ∆u(Bsx )ds 2 0 which is therefore a martingale. By optional stopping theorem we have that " # Z T ∧N 1 x u(x) = E u(BT ∧N ) − ∆u(Bs )ds 2 0 "Z # T ∧N x x = E[u(BT ∧N )] + E g(Bs )ds 0 Since u is twice differentiable on D we have that u is bounded hence by DCT we get that lim E[u(BTx ∧N )] = 0 N →∞ since u(x) = 0 in the boundary. Similarly Z T ∧N g(Bsx )ds ≤ ||g||∞ T 0 and E[T ] < ∞ so DCT gives us that "Z lim E N →∞ # T ∧N g(Bsx )ds 0 "Z =E 0 So indeed we get the desired result. 29 T # g(Bsx )ds We can extend the Dirichlet and Poisson problems to hold for u ∈ C 2 (D) ∩ C(D) by using the case we have proven on Dε = {x : d(x, Dc ) < ε} and taking the limit as ε → 0. Theorem 5.4. If u ∈ C 1,2 ([0, ∞) × Rd ) is of at most exponential growth and is a solution to ( ∂u 1 t > 0, x ∈ Rd ∂t = 2 ∆u u(0, x) = f (x) x ∈ Rd then u(x, t) = E[f (Btx )] |x−y|2 e− 2t dy = f (x) (2π)d/2 Rd Z Proof. If we fix t > 0 then u(Bsx , t − s) − Z s − 0 ∂u 1 + ∆u (Brx , t − r)dr ∂t 2 is a martingale by lots of Brownian motion, moreover since is null so u(Bsx , t − s) is a martingale. By optional stopping theorem we have that ∂u ∂t = 21 ∆u we have that the integral term u(x, t) = E[u(Btx , 0)] = E[f (Btx )] Theorem 5.5. For D ⊆ Rd ff u ∈ C 1,2 ([0, ∞) × D) to ∂u 1 ∂t = 2 ∆u u(0, x) = f (x) u(t, x) = g(x) is of at most exponential growth and is a solution t > 0, x ∈ D x∈D x ∈ ∂D, t > 0 then u(x, t) = E[g(BTx )IT ≤t + f (Btx )IT >t ] Proof. If we fix t > 0 then u(Bsx , t Z s − s) − 0 ∂u 1 − + ∆u (Brx , t − r)dr ∂t 2 1 is a martingale by lots of Brownian motion, moreover since ∂u ∂t = 2 ∆u we have that the integral term x is null so u(Bs , t − s) is a martingale. T ∧ t is a bounded stopping time so setting s = 0, optional stopping theorem gives us that u(x, t) = E[u(BTx ∧t , t − (T ∧ t))] = E[χT <t g(BTx ) + χT ≥t f (Btx )] Definition 5.2. Convex: φ : R → R is convex if φ(x) = sup{L(x) : L ≤ φ, L linear} Proposition 5.2. Jensen’s Inequality: If E[|X|] < ∞, φ : R → R is convex then both E[φ(X)] ≥ φ(E[X]) E[φ(X)|F] ≥ φ(E[X|F]) hold a.s. 30 Corollary 5.10. If {Mt }t≥0 is a martingale and φ : R → R is convex then E[φ(Mt )|Fs ] ≥ φ(E[Mt |Fs ]) = φ(Ms ) Lemma 5.1. Fatou: If {Xn }∞ n=1 is a sequence of positive random variables then E[lim inf Xn ] ≤ lim inf E[Xn ] n n Lemma 5.2. If {Mt }t≥0 is a martingale with continuous paths, φ : R → R+ convex and T ≤ K is a bounded stopping time then E[φ(Mt )] ≤ E[φ(MK )] Proof. We shall split the proof into then discrete and continuous cases: • Case 1) Suppose T ∈ {tr }m r=1 are the discrete times T can take ordered such that tr < tr+1 ≤ K E[φ(MT )] = ≤ m X r=1 m X E[φ(Mtr )χ{T =tr } E[φ(MK )χ{T =tr } r=1 = E[φ(MK )] • Case 2) Suppose T is a general bounded stopping time then find a sequence of discrete bounded stopping times {Tn }∞ n=1 such that Tn ≥ Tn+1 ≥ T and limn→∞ Tn = T From case 1 we have that for any n E[φ(MTn )] ≤ E[φ(MK )] and by Fatou’s lemma we have that E[φ(MT )] = E[lim inf φ(MTn )] ≤ lim inf E[φ(MTn )] ≤ E[φ(MK )] n n Theorem 5.6. Optional Stopping Theorem (version 2): If {Mt }t≥0 is a martingale on {Ft }t≥0 with continuous paths and T a bounded stopping time then E[MT ] = E[M0 ] Proof. From version 1 we have that then statement holds when {Mt }t≥0 is bounded or when T only takes discrete values. Write (x)+ := max{0, x} which is convex and x = (x − L)+ − (−L − x)+ + max{min{x, L}, −L} In particular MT = (MT − L)+ − (−L − MT )+ + max{min{MT , L}, −L} Find a sequence of discrete bounded stopping times {Tn }∞ n=1 such that Tn ≥ Tn+1 ≥ T and limn→∞ Tn = T Fix ε > 0 and choose L sufficiently large such that E[(MTn − L)+ ] ≤ E[(MK − L)+ ] < ε and similarly E[(−MTn − L)+ ] ≤ E[(−MK − L)+ ] < ε 31 We can then choose N sufficiently large such that |E[max{min{MTn , L}, −L}] − E[max{min{MT , L}, −L}]| ≤ ε for any n ≥ N since max{min{MT , L}, −L} is bounded. Moreover E[max{min{x, L}, −L}] = E[MT ] − E[(MT − L)+ ] + E[(−L − MT )+ ] so |E[max{min{MTn , L}, −L}] − E[MT ] − E[(MT − L)+ ] + E[(−L − MT )+ ]| ≤ ε Which gives |E[max{min{MTn , L}, −L}] − E[MT ]| ≤ 3ε Since TN is discrete from version 1 we have that E[M0 ] = E[MTN ] = E[(MTN − L)+ ] − E[(−L − MTN )+ ] + E[max{min{MT , L}, −L}] So |E[M0 ] − E[MT ]| ≤ 5ε Since ε > 0 was chosen arbitrarily we indeed have that the statement holds. Definition 5.3. Regular: If y ∈ ∂D is regular for u : R → R and f : ∂D → R if for any sequence {xn }∞ n=1 ∈ D such that limn→∞ xn = y we have that limn→∞ u(xn ) = f (y) Definition 5.4. Ball Averaging: u : D → R satisfies ball averaging if for any x ∈ D and any ε > 0 s.t. B(x, ε) ⊂ D we have that Z 1 u(x) = u(y)dy VB(x,ε) B(x,ε) where VB(x,ε) is the volume of the ball B(x, ε). Definition 5.5. Sphere Averaging: u : D → R satisfies sphere averaging if for any x ∈ D and any ε > 0 s.t. B(x, ε) ⊂ D we have that Z 1 u(x) = u(y)dy S B(x,ε) ∂B(x,ε) where SB(x,ε) is the surface area of the ball B(x, ε). Lemma 5.3. If u solves the Dirichlet problem then it satisfies sphere averaging. Proof. Let Sε := inf {t ≥ 0 : Btx ∈ ∂B(x, ε)} be the first time the Brownian motion leaves the ball B(x, ε) then Xt = BSxε +t − BSε is a Brownian motion by the strong Markov property. Furthermore let T 0 := inf {t > 0 : Xt + BSε ∈ ∂D} be the time between leaving the ball for the first time and leaving D. We then have that u(x) = E[f (BTx )] = E[f (BSxε + BTx − BSxε )] = E[f (BSxε + XT 0 )] = E[E[f (BSxε + XT 0 )|FSε ]] = E[f (z + XT 0 )|z=BSε ] = E[u(z)|z=BSε ] = E[u(BSxε )] which is the sphere average. 32 Lemma 5.4. If u(x) = E[f (BTx )] ∈ C ∞ (D) with ∆u = 0 on D and every y ∈ ∂D is regular then u solves the Dirichlet problem. Lemma 5.5. If ∃C ⊂ Dc cone with vertex y ∈ ∂D then y is regular. 6 Donsker’s Theorem For {Zi }∞ i=1 i.i.d. random variables with E[Zi ] = 0, V ar[Zi ] = 1 we denote n X Sn := Zi i=1 to be the discrete time Markov chain representing the position of a particle at time n and (N ) Xt SN t := √ N to be the scaled version of Sn . Definition 6.1. E-Valued Random Variable: If (Ω, F, P) is a probability space, (E, d) a metric space and X : (Ω, F, P) is measurable then X is called an E-valued random variable. Lemma 6.1. Let B(C[0, 1]) denote the σ-algebra generated by Borel sets on C[0, 1]. If F0 is the Tn π-system formed by the sets {f ∈ C[0, 1] : i=1 {f (ti ) ∈ Oi }} for open sets {Oi }ni=1 and times {ti }ni=1 then σ(F0 ) = B(C[0, 1]) Definition 6.2. Convergence In Distribution: d If {Xn }∞ n=1 , X are random variables then Xn → X if lim [F (Xn )] = E[F (X)] n→∞ ∀F ∈ Cb . Theorem 6.1. Continuous Mapping Theorem: d ˜ is continuous then If Xn → X on metric space (E, d) and G : (E, d) → (Ẽ, d) d G(Xn ) → G(X) ˜ on (Ẽ, d) Proof. Let F : Ẽ → R be continuous and bounded then F ◦ G is continuous and bounded hence lim E[F (G(Xn ))] = E[F (G(X))] n→∞ so indeed d G(Xn ) → G(X) We denote Disc(f ) := {g : f is discontinuous at g} Theorem 6.2. Entended Continuous Mapping Principle: d If Xn → X on (E, d) and F : E → R is measurable s.t. P(X ∈ Disc(F )) = 0 then d F (Xn ) → F (X) 33 Lemma 6.2. Skorokhod: If Z is a random variable with E[Z] = 0 then ∃T < ∞ stopping time s.t. BT has the same distribution as Z and E[T ] = E[Z 2 ] Proof. Firstly suppose that Z only takes values {a, b} where a < 0 ≤ b Then letting T = Ta,b := inf{t ≥ 0 : Bt ∈ {a, b} and using a previous result we have that b P(BT = a) = b−a −a P(BT = b) = b−a E[T ] = −ab which satisfies the requirements for the lemma. For general T let α < 0 ≤ β be random variables independent of B then the aim is to use Tα,β by choosing an appropriote joint density of α, β : ν(da, db) = P(α ∈ da, β ∈ db) to match Z. Write µ+ (dz) = P(Z ∈ dz) for z ≥ 0 and µ− (dz) = P(Z ∈ dz) for z < 0. For z ≥ 0 we want that µ+ (dz) = P(BTα,β ∈ dz) We have that P(BTα,β ∈ dz) = E[P(BTα,β ∈ dz|σ(α, β))] α χβ∈dz =E − β−α Z Z a = − χb∈dz ν(da, db) b−a Z a = − ν(da, dz) z−a from previous case So choose ν(da, db) = (b − a)µ+ (db)π(da) where Z −aµ+ (dz)π(da) = 1 which gives us that Z P(BTα,β ∈ dz) = −aµ+ (dz)π(da) = µ+ (dz) as required. For z < 0 we want that µ− (dz) = P(BTα,β ∈ dz) We have that P(BTα,β ∈ dz) = E[P(BTα,β ∈ dz|σ(α, β))] β =E χα∈dz β−α Z Z b = χa∈dz ν(da, db) b−a Z b = ν(dz, db) b−z Z = bµ+ (db)π(dz) from previous case 34 By choosing π(dz) = R µ− (dz) xµ+ (dx) we have that Z P(BTα,β ∈ dz) = bµ+ (db) R µ− (dz) xµ+ (dx) = µ− (dz) as required hence we have the distribution ν(da, db) = (b − a)µ+ (db)µ− (da) R xµ+ (dx) as a joint density. By construction we have that P(BTα,β ∈ dz) = P(Z ∈ dz) so it remains to show that 1. Z −a R µ− (da) =1 bµ+ (dx) 2. E[Tα,β ] = E[Z 2 ] 3. Z Z ν(da, db) = 1 Which hold as follows: 1. 0 = E[Z] Z Z = xµ+ (dx) + xµ− (dx) hence Z Z xµ+ (dx) = So indeed Z −a R µ− (da) = bµ+ (dx) 2. Notice that 2 E[Z ] = Z Z −xµ− (dx) Z −aµ− (da) bµ+ (db) = 1 Z 2 x µ+ (dx) + 35 x2 µ− (dx) So we have that E[Tα,β ] = E[−αβ] Z = −abν(da, db) Z Z −ab(b − a)µ+ (db)µ− (da) R = xµ+ (dx) Z Z Z Z 2 2 −ab µ+ (db)µ− (da) a bµ+ (db)µ− (da) R R = + xµ+ (dx) xµ+ (dx) R R 2 R 2 R − aµ− (da) b µ+ (db) a µ− (da) bµ+ (db) R R = + xµ+ (dx) xµ+ (dx) R R 2 R 2 R aµ+ (da) b µ+ (db) a µ− (da) bµ+ (db) R R = + xµ+ (dx) xµ+ (dx) Z Z = b2 µ+ (db) + a2 µ− (da) = E[Z 2 ] 3. Z Z (b − a)µ+ (db)µ− (da) R xµ+ (dx) R RR bµ+ (db)µ− (da) − aµ+ (db)µ− (da) R xµ+ (dx) R R R µ− (da) bµ+ (db) − µ+ (db) aµ− (da) R xµ+ (dx) R R R µ− (da) bµ+ (db) + µ+ (db) aµ+ (da) R xµ+ (dx) Z µ− (da) + µ+ (db) Z Z ν(da, db) = R = R = R = Z = =1 Theorem 6.3. Donsker’s Theorem: If {Zi }∞ i=1 i.i.d. random variables with E[Zi ] = 0, V ar[Zi ] = 1 and Sn := n X Zi i=1 is the discrete time Markov chain representing the position of a particle at time n and (N ) Xt SN t := √ N is the scaled version of Sn then X (N ) converges in distribution to a Brownian motion as N → ∞. Proof. By Skorokhod we can find α, β random variables such that E[Tα,β ] = 1 and BTα,β is equal to Z1 in distribution. Take i.i.d copies {(αi , βi )}∞ i=1 independent of B then let T1 = inf{t ≥ 0 : Bt ∈ {α, β}} and TN +1 = inf{t ≥ TN : Bt − BTN ∈ {αN +1 , βN +1 }} which are stopping times. 36 We can then let BN = √ t N (N ) Bt Furthermore let ΩN,ε = {||X (N ) − B (N ) ||∞ > ε} then we want to show that lim P(ΩN,ε ) = 0 N →∞ since given this for fixed F unifomrly continuous bounded function we have that |E[F (X (N ) )] − E[F (B (N ) )]| ≤ E[|F (X (N ) ) − F (B (N ) )|] ≤ 2||F ||∞ P(ΩN,ε ) + E[|F (X (N ) ) − F (B (N ) )|χΩN,ε ] So for fixed η > 0 by uniform continuity we can choose ε > 0 small enough such that E[|F (X (N ) ) − F (B (N ) )|χΩN,ε ] ≤ η and N large enough such that 2||F ||∞ P(ΩN,ε ) ≤ η So indeed we shall have convergence in distribution. Hence it remains to show that P(ΩN,ε ) < ε which breaks down into: 1. Since (N ) XK N SK BT (N ) = √ = √ K = B TK N N N we need that (N ) (N ) B TK ≈ B K N N (N ) 2. We also require that Xt (N ) − Bt is small on the intervals t ∈ [K/N, (K + 1)/N ). These indeed hold as follows: 1. T1 , T2 − T1 , T3 − T2 , ... are i.i.d with mean 1 so by the strong law of large number we have that TN /N converges to 1 almost surely. Furthermore maxK≤N (TK 0K)/N converges to 0 almost surely as N → ∞ so writing T − K K ΩN,δ = maxk=1,...,N ≥δ N gives us that limN →∞ P(ΩN,δ ) = 0. (N ) 2. Suppose ||X (N ) − B (N ) ||∞ > ε then ∃t ∈ [0, 1] s.t. |Xt WLOG we can let K ≤ N t < K + 1 then either (N ) |Bt (N ) − Bt | ≥ ε. (N ) − BK/N | ≥ ε or (N ) |Bt (N ) − B(K+1)/N | ≥ ε So we have that (N ) P(||X (N ) − B (N ) ||∞ > ε) ≤ P(ΩN,δ ) + P(|Bs(N ) − Bt (N ) (N ) > ε f or some |s − t| ≤ δ + N −1 ) But P(|Bs − Bt > ε f or some |s − t| ≤ δ + N −1 ) = P(|Bs − Bt > ε f or some |s − t| ≤ 2δ) by choosing N sufficiently large. By choosing δ sufficiently small and using continuity of paths we then have that P(|Bs − Bt > ε f or some |s − t| ≤ 2δ) can be made sufficinetly small then taking the limit N → ∞ we get the desired result. 37 ∞ Lemma 6.3. If {Xn }∞ n=1 , {Yn }n=1 are sequences of random variables on the same probability space such that Xn converges to X in distribution and Yn converges to 0 in distribution then Xn + Yn converges to X in distribution. Proof. It suffices to show that the characteristic functions converge. |E[eiθ(Xn +Yn ) ] − E[eiθX ]| = |E[eiθ(Xn +Yn ) − eiθXn ] + E[eiθXn − eiθX ]| ≤ |E[eiθ(Xn +Yn ) − eiθXn ]| + |E[eiθXn − eiθX ]| lim |E[eiθXn n→∞ iθ(Xn +Yn ) − eiθX ]| = 0 − eiθXn ]| = |E[eiθXn (eiθYn − 1)]| |E[e ≤ E[|eiθXn ||eiθYn − 1|] ≤ E[|eiθYn − 1|] q ≤ E[|eiθYn − 1|2 ] q = 2E[1 − eiθYn ] by Cauchy-Schwarz lim E[1 − eiθYn ] = 0 n→∞ So indeed Xn + Yn converges to X in distribution. Lemma 6.4. If Xn is a sequence of random variables such that limn→∞ E[|Xn |] = 0 then Xn converges to 0 in distribution. Proof. It suffices to show that the characteristic functions converge. Recall that ex ≥ 1 + x so: |E[1 − eiθXn ]| ≤ E[|θX|] → 0 Theorem 6.4. If {Xn }∞ n=1 , X are random variables on the same probability space and (E, d) is a metric space then the following are equivalent: 1. limn→∞ E[F (Xn )] = E[F (X)] ∀F : E → R bounded and continuous. 2. limn→∞ E[F (Xn )] = E[F (X)] ∀F : E → R bounded and uniformly continuous. 3. lim supn→∞ P(Xn ∈ A) ≤ P(X ∈ A) for any A closed. 4. lim inf n→∞ P(Xn ∈ A) ≥ P(X ∈ A) for any A open. 5. limn→∞ P(Xn ∈ A) = P(X ∈ A) whenever P(X ∈ ∂A) = 0. 6. limn→∞ E[F (Xn )] = E[F (X)] for all bounded, measurable F : E → R s.t. P(X ∈ Disc(F )) = 0. Proof. • 1 =⇒ 2) If F is uniformly continuous then it is also continuous hence the claim is immediate. • 2 =⇒ 3) Let ε > 0 and define Fε (x) = 1− 38 d(x, A) ε + which are uniformly continuous, bounded functions such that lim Fε (x) = χA (x) ε→0 Since Fε (x) ≤ 1 by dominated convergence theorem we have that Z lim E[Fε (X)] = χA dP = P(X ∈ A) n→∞ So we have that P(X ∈ A) = lim E[Fε (X)] ε→0 = lim lim E[Fε (Xn )] ε→0 n→∞ ≥ lim sup E[χA (Xn )] n→∞ = lim sup P(Xn ∈ A) n→∞ • 3 =⇒ 4) Let A open then Ac is closed so: P(X ∈ A) = 1 − P(X ∈ Ac ) ≤ 1 − lim sup P(Xn ∈ Ac ) n→∞ = lim inf P(Xn ∈ A) n→∞ • 3, 4 =⇒ 5) If P(X ∈ ∂A) = 0 then P(X ∈ A) = P(X ∈ A) = P(X ∈ Ao ) so we have that lim sup P(Xn ∈ A) ≤ lim sup P(Xn ∈ A) n→∞ n→∞ ≤ P(X ∈ A) = P(X ∈ A) = P(X ∈ Ao ) ≤ lim inf P(Xn ∈ Ao ) n→∞ ≤ lim inf P(Xn ∈ A) n→∞ ≤ lim sup P(Xn ∈ A) n→∞ • 5 =⇒ 6) Notice that 5 is the specific case of 6 where F = χA since Disc(χA ) = ∂A so we have 6 for indicator functions. For general bounded measurable F fix ε > 0 and choose {αi }ni=1 increasing sequence such that |αi − αi+1 | < ε so we can define the ε approximation of F as Fε (x) = n X αi χf −1 (αi ,αi+1 ] (x) i=1 Notice that pα := P(F (X) = α) > 0 for countably many α only since {α : pα > 1/n} has at most n elements. Letting Q = {α : pα > 0} we can choose {αi }ni=1 ∈ / Q then we have that P(F (X) = α) = 0. 39 So if we let Ai = f −1 (αi , αi+1 ] = {x : f (x) ∈ (αi , αi+1 ]} then we have that Disc(χAi ) ⊂ Disc(F ) ∪ {x : F (x) = αi } ∪ {x : F (x) = αi+1 } and P(Disc(F )) = P({x : F (x) = αi }) = P({x : F (x) = αi+1 }) = 0 hence P(X ∈ ∂Ai ) = 0 so 5 holds with f (x) = χAi (x) and so lim E[χAi (Xn )] = E[χAi (X)] n→∞ We therefore get that lim E[Fε (Xn )] = lim n→∞ n→∞ = lim n→∞ n X i=1 n X αi E[χAi (Xn )] αi E[χAi (X)] i=1 = E[Fε (X)] Furthermore ||F − Fε || ≤ ε but ε was chosen arbitrarily hence indeed limn→∞ E[F (Xn )] = E[F (X)] • 6 =⇒ 1) If F is bounded and continuous then F is measurable hence the claim is immediate. 40