エクセンダールSDEもんだい3.14
Let be the filtration generated by the Brownian motion .
We see obviously that is -measurable if it is a random variable as in the assertion. Then we show the converse.
Suppose is -measurable. We now follow the hint: a)-c).
a) Define for . Then we see that a.s. If, for every , we have a sequence of bounded random variables which converges to in , taking an appropriate yields that .
Then let . We obtain
\begin{align*}
P(|h-g_n|>\varepsilon)
&\leq P(|h-h_n|>\varepsilon/2) + \frac{2E[|h_n-g_n|]}{\varepsilon} \\
&\leq P(|h-h_n|>\varepsilon/2) + \frac{2}{n\varepsilon}
\end{align*}
for . Since tends to as in probability, so does in probability. Thus must have an a.s. convergent subsequence with limit .
Then it suffices to construct a sequence of as in the assertion which converges to any bounded in .
b) We now show that
\begin{align*}
\mathcal{F}_t = \bigvee_{n=1}^\infty \mathcal{H}_n =: \mathcal{H} .
\end{align*}
For , take a sequence from converging to . Since and every is -measurable, is also -measurable. Thus, we see that . And then we get the inverse inclusion by definitions of and .
By Corollary C.9, it follows that a.s. and in .
c) By the Doob-Dynkin lemma (Lemma 2.1.2), we have
\begin{align*}
h_n = G_n(B_{t_1},\dots,B_{t_k}),
\end{align*}
where for . Since are bounded, we see that are also bounded.
Let be the set of all continuous functions , which are univariate functions and whose support are compact when regarding as ones .
We now consider the following:
Claim Any bounded function admits a sequence in the subabgebra generated by such that
\begin{align*}
& \sup_{j\geq 1,~ x\in\mathbb{R}^k}|F_j(x)|<\infty, \quad \lim_{j\to\infty} F_j \to G, \quad \text{a.e.}
\end{align*}
We prepare Lemma 1 and 2 before showing Claim.
Lemma 1 There exists a sequence in *1 such that for some independent of , and a.e.
Proof. Let for . Then every is in for under the Lebesgue measure and tends to a.e. as .
For , by density of in , one can take an -convergent sequence of with limit . Moreover, we may assume that
\begin{align*}
\sup_{x\in\mathbb{R}^k}|F^{(N)}_j(x)|\leq K:=\sup_{x\in\mathbb{R}^k}|G(x)|+1 \quad \text{for $j,N\in\mathbb{N}$}.
\end{align*}
Indeed, defining as
\begin{align*}
\tilde{F}^{(N)}_j(x) &=
\begin{cases}
K & ,\text{if $F^{(N)}_j(x)>K$}, \\
-K & ,\text{if $F^{(N)}_j(x)<-K$}, \\
F^{(N)}_j(x) & ,\text{otherwise},
\end{cases}
\end{align*}
leads that and
\begin{align*}
\int_{\mathbb{R}^k} |\tilde{F}^{(N)}_j(x)-G_N(x)|dx
&\leq \int_{\mathbb{R}^k} |F^{(N)}_j(x)-G_N(x)|dx,
\end{align*}
and then it follows that in . Then the same argument as in a) yields the assertion. □
Lemma 2 For , let be a dimensional compact rectangle containing the support of and let be the set of all continuous functions , . Then each admits a sequence of the subalgebra generated by such that . Furthermore, we can extend the domain of each to whole satisfying .
Proof. We refer to Miyajima [1], Theorem 7.65, for the Stone-Wierstrass theorem:
Theorem Let be a compact Hausdorff space and be a subset of . Then, the subabgebra generated by is dense in if and only if the following (1) and (2) hold:
(1) For any , there exists an such that .
(2) For any different , there exists an such that .
The theorem is stated in other books or media as well, e.g. [2-4].
It is clear that each satisfies (1) and (2) in the theorem. Thus, we see that the former holds.
Next, for the latter. We extend each of the domain of by putting the same values on , where , as ones on the boundary of along each axis (Fig. 1).
And then, rewrite as the product of the extended above and , where each is defined as in Fig. 2. Then we see that each of the extended is in the subalgebra generated by , which is defined before Claim, and . □
Proof of Claim. By Lemma 1 and 2, we may assume that and that , by renumbering if necessary. Then we have
\begin{align*}
|\tilde{F}^{(j)}_j(x)-G(x)|
&\leq |F_j(x)-G(x)| + \frac{1}{j} \\
& \to 0, \quad \text{as $j\to\infty$}
\end{align*}
almost everywhere. □
We are now back to showing the exercise. By Claim, there exists, for , a sequence in the subabgebra generated by converging to a.e. By the dominated convergence theorem, we see that as . Then as repeated above, we assume that
\begin{align*}
E[|G_n(B_{t_1},\dots,B_{t_k}) - F^{(n)}_n(B_{t_1},\dots,B_{t_k})|]\leq \frac{1}{n}.
\end{align*}
Then, letting yields
\begin{align*}
E[|h-F^{(n)}_n(B_{t_1},\dots,B_{t_k})|]
&\leq E[|h-h_n|] + \frac{1}{n} \\
&\quad \to 0.
\end{align*}
As mentioned in a), we constructed a desired convergent sequence. We complete the proof.
References
[1] Miyajima, S., Functional Analysis, Yokohama Publishers, 2007. In Japanese.
[2] Rudin, W., Functional Analysis, McGraw-Hill, 1991.
[3] Yosida, K., Functional Analysis, Springer, 1980.
[4] Wikipedia, Stone–Weierstrass theorem
Hmm, too heavy exercise...x(
*1:It denotes the space of all continuous functions with compact support.
2月27日のしゅくだい
(b)を示せば、とすることで(a)が得られる。よって(a)のみを示す。
中点""でのユークリッド内積を表すとし、とおく。
\begin{align*}
& \partial_t f(t,x) = cf(t,x), \\
& \nabla_x f(t,x) = f(t,x)\alpha, \\
& \partial_{x_ix_j}f(t,x) = f(t,x)\alpha_i\alpha_j
\end{align*}
となる。
であり、伊藤の公式より、
\begin{align*}
dX_t
&= \partial_tf(t,B_t)dt + \nabla_xf(t,B_t)\cdot dB_t + \frac{1}{2}\sum_{i,j=1}^n \partial_{x_ix_j}f(t,B_t)dB^i_tdB^j_t \\
&= c X_t dt + X_t \alpha\cdot dB_t + \frac{1}{2}|\alpha|^2X_tdt \\
&= \left( c+\frac{1}{2}|\alpha|^2 \right)X_tdt + X_t \alpha\cdot dB_t
\end{align*}
を得る。
Lévy processのジャンプはいつまでも起き続ける
The following might be nontrivial so I will write down a proof.
Claim
Assume that the filtration satisfies the usual conditions. Let be a Lévy process (with càdlàg sample paths) adapted to and be a Borel set in such that . Define the random times
\begin{align*}
T^0_\Lambda &= 0, \\
T^{n}_\Lambda &= \inf\{t>T^{n-1}_\Lambda;\Delta X_t\in\Lambda\}, \quad n=1,2,\dots
\end{align*}
, where . Then the followings hold:
(i) a.s. for
(ii) Every is an -stopping time.
(iii) a.s.
Proof.
(i) Since , holds. Then implies .
By right continuity of , there is a such that
\begin{align*}
\left|X_t-X_{T^{n-1}_\Lambda}\right| \leq \frac{d}{6}
\end{align*}
for all .
For any , taking a sufficiently small satisfying yields, by càdlàgness of ,
\begin{align*}
\left| X_{t-{\delta_2}}-X_{T^{n-1}_\Lambda} \right|,|X_{t-{\delta_2}}-X_{t-}| \leq \frac{d}{6}.
\end{align*}
Thus, it holds that
\begin{align*}
\left|\Delta X_t\right|
&\leq |X_t-X_{T^{n-1}_\Lambda}| + |X_{t-{\delta_2}}-X_{T^{n-1}_\Lambda}| + |X_{t-{\delta_2}}-X_{t-}| \\
&\leq \frac{d}{6}+\frac{d}{6}+\frac{d}{6} = \frac{d}{2} < d.
\end{align*}
Then we obtain .
(ii) We use the result from Karatzas-Shreve; Proposition 2.26.
It holds for , since is -adapted,
\begin{align*}
\{ T^n_{\Lambda} \leq t \}
&= \bigcup_{k_1,\dots,k_n\in\mathbb{N}} \bigcap_{j=1}^n \left( \{T_{k_j}\leq t\} \cap \{ \Delta X_{T_{k_j}}\in\Lambda \} \right) \\
&\in \mathcal{F}_t,
\end{align*}
where, by the proposition above, is the sequence of stopping times exhausting the jumps of . Thus we see all are stopping times.
(iii) Assume that there is a such that for all . Then, is increasing and bounded. Thus, exists.
By càdlàgness of , it follows . Since is left continuous, we get .
Hence, as . However, . Those are contradiction.
2月13日のしゅくだい4
、をの部分-加法族のとき、
\begin{align*}
E[(E[X|\mathcal{G}])^2] \leq E[X^2] \tag{1}
\end{align*}
しょうめい:
は凸関数なので、イェンセンの不等式より、
\begin{align*}
(E[X|\mathcal{H}])^2 \leq E[X^2|\mathcal{H}]
\end{align*}
である。
よって、両辺期待値をとれば、(1)を得る。□
補題6.1.1を見よともあります。
ヒルベルト空間において、可測な確率変数のなす部分空間をとする。
補題6.1.1より、]はへの直交射影なので、
\begin{align*}
E[(E[X|\mathcal{G}])^2]
&= \Vert P_{\mathcal{N}}X \Vert_{L^2}^2 \\
&\leq \Vert P_{\mathcal{N}} \Vert^2 \Vert X \Vert_{L^2}^2 \\
&\leq \Vert X \Vert_{L^2}^2 = E[X^2]
\end{align*}
となる。
2月13日のしゅくだい3
(i) ならば、とは独立
(ii) は定常
(iii) すべてのに対し、]
を満たす実数値確率過程で連続なものはなるものを除いて存在しない。
しょうめい:
Step 1) に対し、とする。
このとき、各に対して、は連続であるので、これととの合成であるは(i),(ii)を満たす。
(ヒントの通り計算する。)
\begin{align*}
E[(W^N_t - W^N_s)^2]
&= E[(W^N_t)^2] - 2E[W^N_tW^N_s] + E[(W^N_s)^2]\\
&= 2E[(W^N_0)^2 - (E[W^N_0])^2] \\
&= 2E[(W^N_u - E[W^N_0])^2]
\end{align*}
が任意ので成り立つ。
今、が連続過程であると仮定する。
すると、も連続であり、したがってとすると、優収束定理(有界収束定理)より
\begin{align*}
E[(W^N_u - E[W^N_0])^2] = 0
\end{align*}
を得る。
よって、
\begin{align*}
W^N_u = E[W^N_0] \tag{1}
\end{align*}
である。
よって、は定数過程である。
Step 2) 任意のに対し(1)が成り立つことよりなので、
\begin{align*}
&\{W_0>N\} = \{W_u>N\} \\
&\{W_0<-N\} = \{W_u<-N\} \\
&\{-N\leq W_0\leq N\} = \{-N\leq W_0\leq N\}
\end{align*}
が分かる。
このとき、任意のに対して、次の①~③のいずれか1つのみが成り立つ:
①
②
③
∵) ①については、とおくと、すでに述べたようにで、(i)よりとは独立であること、そして(ii)より、
\begin{align*}
P(A_0)
&= P(A_0\cap A_u) \\
&= P(A_0)P(A_u) \\
&= P(A_0)^2
\end{align*}
となり、を得られる。②、③も同様。①~③の事象は互いに素なので、いずれか1つのみが成り立つ。
Step 3) 任意のに対して、が成り立つことを示す。
あるで①が成り立つとすると、となり、に反する。
②も同様。
よって、任意のに対し、である。
Step 3)より、特にとしておくと、
\begin{align*}
W_u
&= W^N_u \\
&= E[W^N_0] \\
&= E[W_0] = 0
\end{align*}
となり、を得る。このは(i)~(iii)を満たす。□
2月13日のしゅくだい2
\begin{align*}
N_t = B_t^3-3tB_t
\end{align*}
がマルチンゲールである。
しょうめい:
をブラウン運動によって生成されるフィルトレーションとする。
ブラウン運動はガウス分布に従うので、は可積分である。
とする。
\begin{align}
N_t
&= (B_t-B_s + B_s)^3 - 3tB_t \\
&= (B_t-B_s)^3 + 3B_s(B_t-B_s)^2 + 3B_s^2(B_t-B_s) + B_s^3 - 3tB_t
\end{align}
である。
第1項の条件付き期待値は、
\begin{align*}
E[(B_t-B_s)^3|\mathcal{F}_s]
&= E[(B_t-B_s)^3] \\
&= \frac{1}{\sqrt{2\pi (t-s)}} \int x^3 e^{-\frac{x^2}{2(t-s)}} dx \\
&= 0,
\end{align*}
第2項は
\begin{align*}
E[B_s(B_t-B_s)^2|\mathcal{F}_s]
&= B_sE[(B_t-B_s)^2|\mathcal{F_s}] \\
&= B_sE[(B_t-B_s)^2] \\
&= (t-s)B_s,
\end{align*}
第3項は
\begin{align*}
E[B_s^2(B_t-B_s)|\mathcal{F}_s]
&= B_s^2 E[B_t-B_s] \\
&= 0
\end{align*}
である。
よって、がマルチンゲールになることと合わせると
\begin{align*}
E[N_t|\mathcal{F}_s]
&= 3(t-s)B_s + B_s^3 -3tB_s \\
&= B_s^3 - 3sB_s \\
&= N_s
\end{align*}
となる。□