This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
0, Rµ is a bounded operator from H−d/2−ǫ into Hd/2+2ǫ+2 by (11.38). By m/2 (11.42) this implies that Rµ also bounded operator from H−d/2−ǫ into W2, p,Φ . Hence, we obtain the second inequality in (11.47). ⊔ ⊓ Lemma 11.11 was proved in Kotelenez (1995b, Lemma 4.12) for the case Φ ≡ 1. Its proof can be generalized to the case Φ ∈ {1, ̟ }. However, we provide a different proof, based on the preceding calculations. Lemma 11.11. Suppose f, g ∈ H0,Φ . Set for m ≥ m(0) and D > 0 m/2
gµ := (Rµ,D − I )g + g. Then, lim
µ→∞
×
µm/2 m 2
−1 !
∞ 0
−µu m/2−1
du e
u
dr Φ(r )
dq G(u, r − q) f (q) − f (r )gµ (r ) = 0.
(11.50)
11.1 Basic Estimates and State Spaces
Proof. ≤
245
⎫ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎬
dr Φ(r )dq G(u, r − q) f (q) − f (r )gµ (r )
9
×
dr Φ(r )
9
dr Φ(r )
dq G(u, r − q) f (q) − f (r )
dq G(u, r − q)gµ (r )
2
2
√ G(u, q) G(u, q)) ⎪ ⎪ ⎪ ⎪ 9 ⎪ ⎪ 2 ⎪ ⎪ ⎪ ⎪ ⎪ ≤ cΦ dr Φ(r ) dq G(u, r − q) f (q) − f (r ) g0,Φ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ m/2 ⎪ ⎪ (since dq G(u, r − q) = 1 and Rµ,D is a bounded operator on H0,Φ ) ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎭ −→ 0, as u −→ ∞ by (11.15). (by the Cauchy–Schwarz inequality, employing G(u, q) =
√
(11.51) We next change variables v := µu, whence ⎫ ∞ ⎪ µm/2 m/2−1 −µu ⎪ dr Φ(r ) dq G(u, r − q) f (q) − f (r )gµ (r ) ⎪ du e u ⎪ ⎪ m ⎬ 0 − 1 ! 2 =
∞
dv e−v v m/2−1
0
dr Φ(r )
dq G
⎪ ⎪ ⎪ v ⎪ , r − q  f (q) − f (r )gµ (r ). ⎪ ⎭ µ (11.52)
By (11.51) for any v > 0 and r, q ∈ Rd the integrand on the righthand side of (11.52) tends to 0, as µ → ∞. It remains to find an upper bound which is integrable with respect to dv e−v v m/2−1 and independent of µ. 2 dr Φ(r ) dq G(u, r − q) f (q) − f (r ) ≤
2 f 20,Φ
+2
dr Φ(r )
dq G(u, r − q) f (q)
2
.
As, before, by the Cauchy–Schwarz inequality and the fact that dq G(u, r − q) = 1 2 dr Φ(r ) dq G(u, r − q) f (q) ≤ dr Φ(r ) dq G(u, r − q) f (q)2 .
246
11 Proof of Smoothness, Integrability, and Itˆo’s Formula
If Φ ≡ 1 we are done. If Φ = ̟ we employ (15.47) and obtain
dq ̟ (r )̟ −1 (q)̟ (q)G(u, r − q) f (q)2 γ 2 γ ≤2 dq dr ̟ (q)G(u, r − q)(1 + r − q )  f (q)2 ≤ cγ 1 + u γ f 20,̟ dr
by the integrability of the moments of the normal distribution. Thus, ⎫ v dr ̟ (r ) dq G( , r − q) f (q) − f (r )gµ (r ) ⎪ ⎪ ⎪ ⎪ µ ⎪ ⎪ ⎪ ⎬ 9 γ v ⎪ ≤ cγ ,Φ f 0,̟ g0,̟ 1+ ⎪ ⎪ µ ⎪ ⎪ ⎪ ⎪ ⎭ γ ≤ cγ ,̟ (1 + v ) f 0,̟ g0,̟ for µ ≥ 1.
(11.53)
The righthand side in the last inequality of (11.53) is integrable with respect to dv e−v v m/2−1 . ⊔ ⊓ Lemma 11.12. Let g ∈ Cbm (Rd ; R), m ∈ N. For f ∈ H0 let f g define pointwise multiplication. Then the multiplication operator on H0 , defined by f %−→ f g can be extended to a bounded operator on H−m . Proof. If ϕm ≤ 1, then gϕm ≤ c, where c is the norm of g in Cbm (Rd ; R). Let f ∈ H0 . f g−m = =
sup {ϕm ≤1}
sup {ϕm ≤1}
& f g, ϕ' & f, gϕ' ≤
sup {ψm ≤c}
& f, ψ' ≤ c f −m .
Since H0 is densely and continuously imbedded in H−m , the preceding estimates ⊔ ⊓ imply the extendibility of the multiplication operator g onto H−m .
11.2 Proof of Smoothness of (8.25) and (8.73) (i) Assume for notational simplicity s = 0. Further, it follows from the estimates of the martingale part driven by space–time white noise that we may, without loss of generality, assume σ ⊥ (r, µ, t) ≡ 0. Let m ≥ 0 such that m = m. For 0 ≤ n ≤ m we set
11.2 Proof of Smoothness of (8.25) and (8.73)
247
Q nµ := ∂ n Rµm/2 , Yµ,n (t) := Q nµ Y (t),
(11.54)
where Y (t) is the solution of (8.25) and m has been defined in (11.43). Since Y (·) ∈ C([0, ∞); M f ) a.s. we also have by (11.41) for any ǫ > 0 Y (·) ∈ C([0, ∞); H−d/2−ǫ ). Therefore, for any β ∈ R, Rµβ/2 Y (·) ∈ C([0, ∞); H−d/2−ǫ+β ). Further, let n be a multiindex. By (11.21), (11.24), and the equivalence of the norms  f m and f m for m ∈ N, we easily see that ∂ n Rµm/2 H−α ⊂ H−α+m−n .
(11.55)
Set η := m − m − 2 − d ∈ (0, 2], and choose δ > 0 small enough so that ǫ :=
η−δ > 0. 2
Further, assume, in what follows, n ≤ m. Then, with α = d/2 + δ we obtain from (11.41), (11.44), and (11.55) that with probability 1 for p ≥ 2 Yµ,n (·) ∈ C([0, ∞); H2+d/2+ǫ ) ⊂ C([0, ∞); W2, p,Φ ).
(11.56)
As a result, the following equation holds in Hd/2+ǫ ⊂ H0 ∩ W0, p,Φ ⊂ H0 : Yµ,n (t) = Yµ,n (0) +
−
t 0
▽·
0
t
d 1
2 Dkℓ (s)∂kℓ Yµ,n (s)ds 2 k,ℓ=1
Q nµ (Y (s)F(s))ds
−
0
t
▽ · Q nµ (Y (s)dm(s)). (11.57)
Recall that Y (t, ω) ∈ M f for all ω. Expression (11.41), in addition to mass conservation, implies for ǫ > 0 Y (t, ω)−d/2−ǫ ≤ cγ f (Y )(t, ω)) = cγ f (Y (0, ω)),
(11.58)
248
11 Proof of Smoothness, Integrability, and Itˆo’s Formula
where we may choose ǫ > 0 small. By (11.55) ∂ n Rµm/2 ∈ L(H−α , H−α+m−n ), whence for small ǫ > 0 ∂ n Rµm/2 ∈ L(H−α , H−α+d+2+2ǫ ). Choosing α = d/2 + ǫ, this imlies
∂ n Rµm/2 ∈ L(H−d/2−ǫ , Hd/2+2+ǫ ).
Hence, by (11.41) and (11.42), there is a cµ < ∞ such that uniformly in (t, ω), 0 ≤ n ≤ m, p ≥ 2, i ≤ 2, and any n¯ ∈ N ⎫ n¯ p n( ¯ p−2) n¯ p ⎪ Yµ,n (t, ω)2+d/2+ǫ Yµ,n (t, ω)i, p,Φ ≤ c2 ⎬ ¯ p−2) n¯ p n(
≤ cµ c2
n¯ p
n¯ p
n( ¯ p−2) n¯ p γ f (Y (0, ω))
Y (t, ω) −d/2−ǫ ≤ cµ cc2
⎪ n¯ p n¯ p =: cµ cω < ∞, ⎭
(11.59)
where we used (11.58) in the last inequality. Thus, n¯ p
( p−2)n¯
max sup Yµ,n (t, ω)i, p,Φ ≤ c2
0≤n≤m t≥s
≤ (cµ cω )
n¯ p
max sup Yµ,n (t, ω)2+d/2+ǫ
0≤n≤m t≥s n¯ p
< ∞.
(11.60) n¯ p Yµ,n (t, ω)i, p,Φ
We conclude that, with arbitrarily large probability, is integrable with respect to dt ⊗ dP, i = 0, 1, 2. By mass conservation, this probability depends only on the initial mass distribution. We will initially analyze the case n¯ = 1. The preceding considerations allow us to apply Itˆo’s formula for p ≥ 2 and n ≤ m: ⎫ p p Yµ,n (t)0, p,Φ = Yµ,n (0)0, p,Φ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ d ⎪ t
⎪ p ⎪ p−1 2 ⎪ ⎪ &Yµ,n (s), ∂kℓ Yµ,n (s)'0,Φ Dkℓ (s)ds + ⎪ ⎪ 2 ⎪ ⎪ k,ℓ=1 0 ⎪ ⎪ t ⎪ ⎪ ⎪ p−1 n ⎪ ⎪ −p &Yµ,n (s), ▽ · Q µ (Y (s)F(s))'0,Φ ds ⎪ ⎪ ⎪ 0 ⎪ ⎪ ⎪ t ⎪ ⎪ ⎬ p−1 n −p &Yµ,n (s), ▽ · Q µ (Y (s)dm(s))'0,Φ 0 ⎪ ⎪ ⎪ ⎪ t ⎪ ⎪ p( p − 1) p−2 n ⎪ &Yµ,n (s), [▽ · Q µ (Y (s)dm(s))]'0,Φ ⎪ + ⎪ ⎪ 2 ⎪ 0 ⎪ ⎪ ⎪ ⎪ 5 ⎪
⎪ p ⎪ ⎪ = : Yµ,n (0)0, p,Φ + Ai,n,Y (t) ⎪ ⎪ ⎪ ⎪ i=2 ⎪ ⎪ t
⎪ ⎪ p = Yµ,n (0)0, p,Φ + ai,n,Y (s)ds + A4,n,Y (t), ⎪ ⎪ ⎪ ⎭ 0 i∈{2,3,5}
(11.61)
11.2 Proof of Smoothness of (8.25) and (8.73)
249
where the ai,n,Y (s) denote the integrands in the deterministic integrals of (11.61). (ii) We first decompose a2,n,Y (t). Integrating by parts, we obtain p−1 p p Yµ,n (s, r )(∂ℓ Yµ,n (s, r ))∂k Φ(r )dr = ∂ℓ (Yµ,n (s, r ))∂k Φ(r )dr (11.62) p 2 = − Yµ,n (s, r )∂kℓ Φ(r )dr. Therefore,
⎫ ⎪ ⎪ ⎪ d ⎪ ⎪ p( p − 1)
⎪ p−2 ⎪ =− (Yµ,n (s, r )) (∂k Yµ,n (s, r ))(∂ℓ Yµ,n (s, r ))Φ(r )dr Dkℓ (s) ⎪ ⎪ ⎬ 2 k,ℓ=1 d ⎪ 1
⎪ 2 ⎪ ⎪ + (Yµ,n (s, r )) p ∂kℓ Φ(r )dr Dkℓ (s) ⎪ ⎪ 2 ⎪ ⎪ k,ℓ=1 ⎪ ⎭ =: a2,n,1,Y (s) + a2,n,2,Y (s). (11.63) Next, recalling the homogeneity assumption (8.33), a2,n,Y (s)
a5,n,Y (s) =
∞ ∞ d µm p( p − 1)
du dv e−µ(u+v) u m/2−1 v m/2−1 (Yµ,n (s, r )) p−2 m 2 0 2 [( − 1)!] 0 2 k,ℓ=1 n n × ∂k,r ∂r G(u, r − q)∂ℓ,r ∂r G(v, r − q)Y ˜ (s, q)Y (s, q) ˜ D˜ k,ℓ (s, q − q)Φ(r ˜ )dq dq˜ dr
By Chap. 8, ((8.33) and (8.34)), Dkℓ (s) := D˜ kℓ (0, s).
Set
Dˆ kℓ (s, r − q) ˜ := − D˜ kℓ (s, r − q) ˜ + Dkℓ (s),
(11.64)
where, by the symmetry assumption on D˜ kℓ (s, q˜ − r ) we have Dˆ kℓ (s, r − q) ˜ = Dˆ kℓ (s, q˜ − r ).
Altogether, we obtain a5,n,Y (s) =
∞ ∞ d p( p − 1)
µm du dv e−µ(u+v) u m/2−1 v m/2−1 (Yµ,n (s, r )) p−2 m 2 0 0 2 [( − 1)!] 2 k,ℓ=1 n n × ∂k,r ∂r G(u, r − q)∂ℓ,r ∂r G(v, r − q)Y ˜ (s, dq)Y (s, dq)D ˜ kℓ (s)Φ(r )dr
⎫ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎬
⎪ ∞ ∞ d ⎪ ⎪ µm p( p − 1)
⎪ du dv e−µ(u+v) u m/2−1 v m/2−1 ⎪ (Yµ,n (s, r )) p−2 m ⎪ 2 ⎪ 2 [( 2 − 1)!] 0 0 ⎪ k,ℓ=1 ⎪ ⎪ ⎪ ⎪ n n ˆ ⎪ × ∂k,r ∂r G(u, r − q)∂ℓ,r ∂r G(v, r − q)Y ˜ (s, dq)Y (s, dq) ˜ Dk,ℓ (s, q − q)Φ(r ˜ )dr ⎪ ⎪ −
=: a5,n,1,Y (s) + a5,n,2,Y (s).
⎪ ⎪ ⎪ ⎪ ⎭
(11.65)
250
11 Proof of Smoothness, Integrability, and Itˆo’s Formula
Apparently, a2,n,1,Y (s) + a5,n,1,Y (s) ≡ 0.
(11.66)
(iii) Y (s) ∈ H−d/2−ǫ ∀ǫ > 0. The highest derivative involved in a5,n,2,Y (s) is of the order of n + 1, and after one integration by parts with respect to a scalar coordinate the order can increase to n +2. Applied to Y (s) (in the generalized sense) this would map Y (s) into H−d/2−ǫ−n−2 ⊂ H−m+d/2+ǫ for all small ǫ > 0, since m > d + m + 2 ≥ d + n + 2. Recall from (8.40) the abbreviation for Lkℓ,n (s), where n ≥ 0 with n ≤ m + 1. We have Lkℓ,n (s) := ∂ n D˜ kℓ (s, r )r =0 = −∂ n Dˆ kℓ (s, r )r =0 . (11.67) We expand Dˆ kℓ (s, q − q) ˜ = Dˆ kℓ (s, q − r + r − q) ˜ in the variable r − q, ˜ using Taylor’s formula. Observe that, by symmetry assumption, Lkℓ,n (s) = (−1)n Lkℓ,n (s), whence Lkℓ,n (s) ≡ 0, if n is odd. Further, we recall that Dˆ kℓ (s, 0) = 0. Therefore, ˜ n˜ j + ni , in the following the degree of the polynomials πn˜ j (q − r )πni (r − q), expansion (11.68) is an even number ≥ 2. ⎫ ˜ Dˆ kℓ (s, q − r + r − q) ⎪ ⎪ ⎪ ⎪ ⎪ " # ⎪ ⎪ m+1 ⎪
ˆ ⎪ ni Dkℓ (s, q − r ) ⎪ ⎪ = ∂q−r ˜ πni (r − q) ⎪ ⎪ i! ⎪ ⎪ i=0 ni =i ⎪ ⎪
⎪ ⎪ ⎪ ˆ + ˜ ˜ nm+2 (r − q) θkℓ,nm+2 (s, q, r, q)π ⎪ ⎪ ⎪ ⎪ ⎪ nm+2 =m+2 ⎪ ⎪ ⎪ ⎬ m+1 m+1−i
1 Lkℓ,ni +n˜ j (s)πn˜ j (q − r )πni (r − q) 1{n˜ j +ni ≥2} ˜ ⎪ =− i! j! ⎪ ⎪ i=0 ni =i j=0 n˜ j = j ⎪ ⎪
⎪ ⎪ ⎪ ˆ ⎪ θkℓ,nm+2 (s, q, r, q)π ˜ nm+2 (r − q) ˜ + ⎪ ⎪ ⎪ ⎪ nm+2 =m+2 ⎪ ⎪ ⎪ ⎪ m+1 ⎪
⎪ ˜ ⎪ θkℓ,nm+2 −ni (s, q, r )πnm+2 −ni (q − r )πni (r − q) + ˜ .⎪ ⎪ ⎪ ⎪ ⎪ i=0 ni =i nm+2 −ni =m+2−i ⎪ ⎪ ⎪ ⎭ =: Ikℓ (s, q, q, ˜ r ) + I Is,kℓ (q, q, ˜ r ) + I I Is,kℓ (q, q, ˜ r ), (11.68) ˆ ˜ where by (8.35) the remainder terms θ and θ have derivatives which are bounded uniformly in all parameters. By assumption (8.35), for any T > 0
ess sup sup [Lkℓ,n (s, ω) + θˆkℓ,n (s, ·, r, q, ˜ ω)1 0≤n≤(m+1)
ω 0≤s≤T,q,r,q˜
+θ˜kℓ,n (s, q, ·, ω)1 ] < ∞.
(11.69)
11.2 Proof of Smoothness of (8.25) and (8.73)
251
We replace Dˆ k,ℓ (s, q − p) in a5,n,2,Y (s) with terms from the expansion (11.68). If a term from that expansion does not depend on q˜ (or on q), we can integrate by parts to simplify the resulting expression. Let us assume that, without loss of generality, the term does not depend on q. ˜ Note that for any Z ∈ M f m
µ2 ( m2 − 1)!
∞
dv e−µv v m/2−1
0
∂ℓ,r ∂rn G(v, r − q)Z ˜ (dq) ˜ = ∂ℓ,r ∂ n Rµm/2 Z (r ).
Thus,17
∂
n
=
Rµm/2 Z
(r )
p−2
µm/2 m 2
−1 !
∞
−µv m/2−1
dv e
v
0
˜ (dq) ˜ ∂ℓ,r ∂rn G(v, r − q)Z
1 ∂ℓ,r ((∂ n Rµm/2 Z )(r )) p−1 . p−1
(11.70)
Step 1: n = 0. Let us first analyze Ikℓ (s, q, q, ˜ r ). Replacing Dˆ k,ℓ (s, q − p) with a term from Ikℓ (q, q, ˜ r ) in a5,0,2,Y (s) yields d µm p( p − 1)
− (Yµ,n (s, r )) p−2 m 2 [( 2 − 1)!]2 k,ℓ=1
× ×
where
0
∞
∞
du dv e−µ(u+v) u (m/2)−1 v (m/2)−1
0
(∂k,r G(u, r − q))(∂ℓ,r G(v, r − q)) ˜
⎫ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎬
⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ˜ )dr ⎪ ×Y (s, dq)Y (s, dq)π ˜ n˜ j (q − r )πni (r − q)Φ(r ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ 1 ⎪ ⎭ ×1{n˜ j +ni ≥2} Lkℓ,ni +n˜ j (s), i! j!
(11.71)
n˜ j := nm+2 − ni . Let us first consider the case ni  = 0 and employ (11.70) with Z := Y (s). Integrating by parts in a5,0,2,Y (s) with πn˜ j (q − r ) 1j! Lkℓ,n˜ j (s) replacing Dˆ k,ℓ (s, q − p), we obtain
17
We provide the simple formula (11.70) in a more abstract formulation which allows us to employ it later to smoothed versions of Y (s) (cf. (11.92)).
252
11 Proof of Smoothness, Integrability, and Itˆo’s Formula
⎫ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ k,ℓ=1 ⎪ ⎪ ⎪ ∞ ∞ ⎪ ⎪ ⎪ −µ(u+v) (m/2)−1 (m/2)−1 ⎪ ⎪ × du dv e u v ⎪ ⎪ ⎪ 0 0 ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ × (∂k,r G(u, r − q))(∂ℓ,r G(v, r − q))Y ˜ (s, dq)Y (s, dq)π ˜ n˜ j (q − r )Φ(r )dr ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ d ∞ m/2 ⎪
p( p − 1) ⎪ p−2 µ −µu (m/2)−1 ⎪ ⎪ =− (Yµ,0 (s, r )) du e u ⎪ m ⎪ 2 ⎪ ( − 1)! 0 ⎪ 2 k,ℓ=1 ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ × (∂k,r G(u, r − q))Y (s, dq)πn˜ j (q − r )(∂ℓ,r Yµ,0 )(s, r )Φ(r )dr ⎪ ⎪ ⎪ ⎪ ⎪ ⎬ d ∞ ⎪ µm/2 p
2 ⎪ (Yµ,0 (s, r )) p−1 m = G(u, r − q)) ⎪ du e−µu u (m/2)−1 (∂kℓ,r ⎪ ⎪ 2 ⎪ ( 2 − 1)! 0 ⎪ k,ℓ=1 ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ×Y (s, dq)πn˜ j (q − r )Φ(r )dr ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ d ∞ m/2 ⎪
p ⎪ p−1 µ −µu (m/2)−1 ⎪ ⎪ G(u, r − q)) du e u (Yµ,0 (s, r )) (∂ + k,r ⎪ m ⎪ 2 ⎪ − 1)! ( 0 ⎪ 2 k,ℓ=1 ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ×Y (s, dq)(∂ℓ,r πn˜ j (q − r ))Φ(r )dr ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ d ⎪ ∞ m/2 ⎪ µ p
⎪ p−1 −µu m/2−1 ⎪ (Yµ,0 (s, r )) (∂ G(u, r −q)) d ue u + ⎪ k,r ⎪ m ⎪ 2 − 1)! ( 0 ⎪ 2 k,ℓ=1 ⎪ ⎪ ⎪ ⎪ ⎭ ×Y (s, dq)πn˜ j (q−r )∂ℓ,r Φ(r )dr. d µm p( p − 1)
(Yµ,0 (s, r )) p−2 m − 2 [( 2 − 1)!]2
(11.72)
The first term on the righthand side of (11.72) can be estimated, employing Lemma 11.10. We set n¯ := 1k + 1ℓ , n¯¯ := n˜ j .
¯¯ = n˜ j  ≥ n, ¯ since ni  = 0 and ni  + n˜ j  ≥ 2 (recalling We note that n that the sum has to be even and ≥2). Employing H¨older’s inequality to separate (Yµ,0 (s, r )) p−1 from the integral operator in addition to (11.47), we obtain that the first term is estimated above by
11.2 Proof of Smoothness of (8.25) and (8.73)
253
⎫ d m/2 ∞ ⎪ µ p ⎪ 2 G(u, r − q)) ⎪ du e−µu u (m/2)−1 (∂kℓ,r (Yµ,0 (s, r )) p−1 m ⎪ ⎪ ⎪ 2 ( 2 − 1)! 0 ⎪ k,ℓ=1 ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ×Y (s, dq)πn˜ j (q − r )Φ(r )dr ⎪ ⎪ ⎪ ⎪ ⎪ ⎬ p
( p−1)/ p
⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎭
≤ Yµ,0 (s)0, p,Φ cΦ,m,k,ℓ,n˜ j ,D Yµ,0 (s)0, p,Φ (assuming µ ≥ Dθ + 1) p
= cΦ,m,k,ℓ,n˜ j ,D Yµ,0 (s)0, p,Φ .
(11.73)
The second and third terms are estimated in the same way, where for the third term we also use (15.46). Consider the case when ni  ≥ 1 and n˜ j  ≥ 1 in Ikℓ (s, q, q, ˜ r ). (The case n˜ j  = 0 obviously leads to the same estimate as the previous case.) We now do not integrate by parts, but otherwise we argue as in the previous step, applying Lemma 11.10 to both factors containing Y (s, dq) and Y (s, dq) ˜ in addition to H¨older’s inequality. Thus, we obtain a similar bound as in (11.73): ⎫ d ⎪ ⎪ p( p − 1) µm p−2 ⎪ ⎪ (Yµ,0 (s, r )) ⎪ m ⎪ 2 2 ⎪ [( − 1)!] ⎪ 2 k,ℓ=1 ⎪ ⎪ ⎪ ⎪ ⎪ ∞ ∞ ⎪ ⎪ ⎪ −µ(u+v) (m/2)−1 (m/2)−1 ⎪ v × du dv e u ⎪ ⎪ ⎪ 0 0 ⎬
⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ p ⎪ ˜ )dr ≤cΦ,m,k,ℓ,n˜ j ,D Yµ,0 (s)0, p,Φ . ⎪ ×Y (s, dq)π ˜ n˜ j (q − r )πni (r − q)Φ(r ⎭
×
(∂k,r G(u, r − q))(∂ℓ,r G(v, r − q))Y ˜ (s, dq)
(11.74)
We now estimate I I Is,kℓ (q, q, ˜ r ). The estimate of I Ikℓ (s, q, q, ˜ r ) is easier. It suffices to restrict ourselves to the case ni  = 0 and follow the pattern in the estimate of Ikℓ (q, q, ˜ r ). The other cases can be handled similarly. We obtain
254
11 Proof of Smoothness, Integrability, and Itˆo’s Formula
d p( p − 1)
µm p−2 (Y (s, r )) µ,0 2 [( m2 − 1)!]2 k,ℓ=1 ∞ ∞ du dv e−µ(u+v) u (m/2)−1 v (m/2)−1 0
0
×
(∂k,r G(u, r − q))(∂ℓ,r G(v, r − q))Y ˜ (s, dq)
×Y (s, dq) ˜ θ˜kℓ,nm+2 (s, q, r )πnm+2 (q−r ) Φ(r )dr
∞ d p
µm/2 du e−µu u (m/2)−1 ≤ (Yµ,0 (s, r )) p−1 m 2 [( 2 − 1)!] 0 k,ℓ=1 2 ˜ × (∂kℓ,r G(u, r − q))Y (s, dq)θkℓ,nm+2 (s, q, r )πnm+2 (q − r )Φ(r )dr
⎫ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎬
∞ ⎪ ⎪ p d µm/2 ⎪ ⎪ + (Yµ,0 (s, r )) p−1 m du e−µu u (m/2)−1 ⎪ ⎪ ⎪ 2 − 1)!] [( 0 ⎪ 2 ⎪ k,ℓ=1 ⎪ ⎪ ⎪ ⎪ ⎪ × (∂k,r G(u, r − q))Y (s, dq)(∂ℓ,r θ˜kℓ,nm+2 (s, q, r ))πnm+2 (q − r )Φ(r )dr ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪
⎪ d ∞ m/2 ⎪ p ⎪ µ ⎪ −µu (m/2)−1 p−1 ⎪ + du e u (Yµ,0 (s, r )) ⎪ m ⎪ 2 ⎪ − 1)!] [( 0 ⎪ 2 k,ℓ=1 ⎪ ⎪ ⎪ ⎪ ⎪ ˜ ⎪ × (∂k,r G(u, r − q))Y (s, dq)θkℓ,nm+2 (s, q, r )(∂ℓ,r πnm+2 (q − r ))Φ(r )dr ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ d ⎪ ∞ m/2 ⎪ p
µ ⎪ −µu p−1 (m/2)−1 ⎪ + du e u (Yµ,0 (s, r )) ⎪ ⎪ m ⎪ 2 − 1)!] [( 0 ⎪ 2 ⎪ k,ℓ=1 ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ × (∂k,r G(u, r − q))Y (s, dq)θ˜kℓ,nm+2 (s, q, r )πnm+2 (q − r )∂ℓ,r Φ(r )dr . ⎭
(11.75)
By the boundedness of ∂ℓ,r θ˜kℓ,nm+2 (s, q, r ) and the fact that ∂ℓ,r πnm+2 (q − r ) has degree ≥ 1 in addition to (15.46), it is sufficient to estimate the first term on the righthand side of (11.75). By (11.8), we obtain
11.2 Proof of Smoothness of (8.25) and (8.73)
255
∞ d p
µm/2 du e−µu u (m/2)−1 (Yµ,0 (s, r )) p−1 m 2 [( − 1)!] 0 2
k,ℓ=1
×
2 G(u, r − q))Y (s, dq)θ˜ (∂kℓ,r kℓ,nm+2 (s, q, r )πnm+2 (q − r )Φ(r )dr
d p
≤ cm,β 2
k,ℓ=1
×
Yµ,0 (s, r ) p−1
µm/2
∞
[( m2 − 1)!] 0
du e−µu u (m+m/2)−1
(s, q, r )Φ(r )dr m+2
G(βu, r − q))
G(βu, r − q))
×Y (s, dq)θ˜kℓ,n
d p
≤ cm,β [ sup θ˜kℓ,n (s, q, r )] 2m+2 2 q,r.s
k,ℓ=1
×
Yµ,0 (s, r ) p−1
×Y (s, dq)Φ(r )dr
µm/2
∞
[( m2 − 1)!] 0
du e−µu u (m+m/2)−1
⎫ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎬
⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ d ⎪ ⎪ β p
p−1 −(m/2)−(m−n/2) m/2 ⎪ = c¯m,β  (µ − △) µ (Y (s))(r )Φ(r )dr  ⎪ Yµ,0 (s, r ) ⎪ ⎪ 2 2 ⎪ ⎪ k,ℓ=1 ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ (by (11.21) and (11.24)) ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ d ⎪
⎪ β −m/2 p p−1 ⎪ m/2 ⎪ (µ − △) Y (s))0, p,Φ Yµ,0 (s)0, p,Φ µ ≤ c¯m,β,Φ ⎪ ⎪ 2 2 ⎪ ⎪ k,ℓ=1 ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ (by H¨older’s inequality and since m ≥ 0) ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ d ⎪
⎪ p p−1 m/2 ⎪ ⎪ ≤ c˜m,β,Φ Yµ,0 (s)0, p,Φ Rµ Y (s)0, p,Φ ⎪ ⎪ 2 ⎪ ⎪ k,ℓ=1 ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ (by (11.24), (11.25), and (11.32)) ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ d ⎪ p
⎪ m/2 p p ⎪ ⎪ (Yµ,0 (s)0, p,Φ + Rµ Y (s)0, p,Φ ), ≤ c˜m,β,Φ ⎪ ⎪ 2 ⎪ ⎭ k,ℓ=1
(11.76)
256
11 Proof of Smoothness, Integrability, and Itˆo’s Formula
where the last step follows from the definition of the norm · m, p,Φ and the following simple inequality18 ab ≤
b p¯ ap + p p¯
for nonnegative numbers a and b, and p > 1 and p¯ :=
Recalling that the functions we obtain
1 i! j! Lkℓ,ni +n˜ j (s) p
p p−1 .
are uniformly bounded in s and ω, m/2
p
a5,0,2,Y (s) ≤ c˜m,Φ ¯ (Yµ,0 (s)0, p,Φ + Rµ Y (s)0, p,Φ ).
(11.77)
(iv) Now we estimate a3,0,Y (s). Taylor’s expansion yields m
1 j (∂ Fk (s, r ))πj (q−r )+ Fk (s, q) = Fk (s, r )+ θ¯k,m+1 (s, q, r )πm+1 (q−r ), j! m=m
j=1
(11.78)
where by assumption (8.35), for any T > 0 sup
ess sup
ω 0≤s≤T,q,
θˆkℓ,m+1 (s, q, ·, ω)] < ∞.
(11.79)
Then, ⎫ ⎪ ⎪ ⎪ −p (Yµ,0 ⎪ m ⎪ ⎪ − 1)! ( ⎪ 2 k=1 ∞ ⎪ ⎪ ⎪ −µu (m/2)−1 ⎪ × (∂k,r G(u, r − q))Fk (s, q) du e u ⎪ ⎪ ⎪ 0 ⎪ ⎪ ⎪ ⎪ ×Y (s, dq)Φ(r )dr ⎪ ⎪ ⎪ ⎪ ⎪ ∞ ⎪ d m/2 ⎪
µ ⎪ p−1 −µu (m/2)−1 ⎪ = −p du e u G(u, r − q)) (Yµ,0 (s, r )) (∂ ⎪ k,r ⎪ m ⎪ ( 2 − 1)! 0 ⎪ k=1 ⎪ ⎪ ⎪ ×Y (s, dq)Fk (s, r )Φ(r )dr ⎪ ⎪ ⎪ ⎪ ⎪ ⎬ d m/2 ∞
d
(s, r )) p−1
µm/2
−1 µ du e−µu u (m/2) (Yµ,0 (s, r )) p−1 m ( − 1)! 0 k=1 ⎡ 2 ⎤ m
1 × (∂k,r G(u, r − q)) ⎣ π (q − r )⎦ Y (s, dq)(∂ j Fk (s, r ))Φ(r )dr j! j j=1 ∞ d
µm/2 p−1 du e−µu u (m/2)−1 −p (Yµ,0 (s, r )) ( m2 − 1)! 0 k=1
× (∂k,r G(u, r − q)) θ¯k,m+1 (s, q, r )πm+1 (q − r )Y (s, dq)Φ(r )dr
−p
=:
18
d
k=1
m=m
[Ik (s) + I Ik (s) + I I Ik (s)].
Cf. (15.2) in Sect. 15.1.2 for a proof.
⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎭
(11.80)
11.2 Proof of Smoothness of (8.25) and (8.73)
257
We estimate Ik (s) for n ≤ m as in (11.72), integrating by parts −p
d
( m2 − 1)!
k=1
× =− =
(Yµ,0 (s, r ))
µm/2
p−1
∞
du e
0
u
⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎬
(∂k,r G(u, r − q))Y (s, dq)Fk (s, r )Φ(r )dr
d
(∂k,r ((Yµ,0 (s, r )) p ))Fk (s, r )Φ(r )dr
k=1
d
(Yµ,0 (s, r )) p (∂k,r Fk (s, r ))Φ(r )dr
k=1
+
d
(Yµ,0 (s, r )) p Fk (s, r )∂k,r Φ(r )dr.
k=1
Applying the assumptions on F in addition to (15.46) yields p
d
k=1
⎫
⎪ −µu m−1 ⎪ ⎪
⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎭
p
(11.81)
(11.82)
Ik (s) ≤ c F,Φ Yµ,0 (s)0, p,Φ .
Repeating the arguments in the estimate of a5,0,2,Y (s) with respect to I Ik (t) and I I Ik (t), we obtain altogether p
p
a3,0,Y (s) ≤ cm, p,Φ (Yµ,0 (s)0, p,Φ + Rµm/2 Y (s)0, p,Φ ).
(11.83)
(vi) The previous steps imply a2,0,Y (s) + a5,0,Y (s) + a3,0,Y (s) p
p
≤ cˆ p,Φ,m (Yµ,0 (s)0, p,Φ + Rµm/2 Y (s)0, p,Φ ) p
p
= cˆ p,Φ (Yµ,0 (s)n, p,Φ + Rµm/2 Y (s)0, p,Φ ).
⎫ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎬ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎭
(11.84)
Integrating both sides of (11.84) against ds from 0 to t we obtain from (11.61) t ⎫ p p p ⎪ ⎪ Yµ,0 (t)0, p,Φ ≤ Yµ,0 (0)0, p,Φ + cˆ p,Φ,m (Yµ,0 (s)0, p,Φ ⎬ 0 (11.85) t ⎪ p p−1 ⎪ +Rµm/2 Y (s) )ds − p &Y (s), ▽ · Rµm/2 (Y (s)dM(s))'0,Φ . ⎭ 0, p,Φ
0
µ,0
We take the mathematical expectation on both sides and may apply the Gronwall lemma in addition to (11.25) (for the norms of the initial condition) and obtain
258
11 Proof of Smoothness, Integrability, and Itˆo’s Formula p
p
sup ERµm/2 Y (t)0, p,Φ ≤ cˆm, p,T,Φ EY (0)0, p,Φ .
(11.86)
0≤t≤T
As µ → ∞, Fatou’s lemma implies the integrability part of (8.38) for n = 0. Step 2: n > 0 – Estimates with Spatially Smooth Processes. The problem for n > 0 is that a generalization of Lemma 11.10 to integral operators from (11.46) to f )(r ) ( K¯ µ f )(r ) := ( K¯ µ,m,n, ¯¯ ¯ n,D ∞ µm/2 m/2−1 −µu e [∂rn¯ ∂ n G(Du, r − q)]πn¯¯ (r − q) f (q)dq du u := m ( 2 − 1)! 0 does not yield good upper bounds, which could be used in estimating a5,n,2,Y (s) and a3,n,Y (s). On the other hand, if Y (·) were already smooth, we could proceed differently. Let us, therefore, replace Y (·) in a5,n,2,Y (s) and a3,n,Y (s) by m/2
Yλ (·) := Rλ
m/2
where λ > Dθ + 1 (cf. 11.54). Then, from (11.65) a5,n,2,Yλ (s)
m/2
Y (·) and Yµ,n,λ (s) := ∂ n Rµ Rλ
Y (s) = Q nµ Yλ (s), ⎫ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎬
d p( p − 1)
µm (Yµ,n,λ (s, r )) p−2 m 2 [( 2 − 1)!]2 ∞ ∞ k,ℓ=1 × du dv e−µ(u+v) u m/2−1 v m/2−1 ⎪ ⎪ 0 0 ⎪ ⎪ ⎪ ⎪ ⎪ n n ⎪ × dq dq˜ ∂k,r ∂r G(u, r − q)∂ℓ,r ∂r G(v, r − q)Y ˜ λ (s, q) ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎭ ˆ ˜ Dk,ℓ (s, q − q)Φ(r ˜ )dr. ×Yλ (s, q)
=−
(11.87)
(11.88)
As in Step 1 we first replace Dˆ k,ℓ (s, q − p) with a term from Ikℓ (q, q, ˜ r ) in a5,n,2,Yλ (s): ⎫ d ⎪ p( p − 1)
µm ⎪ p−2 ⎪ − (Yµ,n,λ (s, r )) ⎪ ⎪ m 2 ⎪ 2 − 1)!] [( ⎪ 2 k,ℓ=1 ⎪ ∞ ∞ ⎪ ⎪ ⎪ −µ(u+v) m/2−1 m/2−1 ⎪ ⎪ v × du dv e u ⎪ ⎪ ⎬ 0 0 × dq dq(∂ ˜ k,r ∂rn G(u, r − q))(∂ℓ,r ∂rn G(v, r − q))Y ˜ λ (s, q) ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ×Yλ (s, q)π ˜ )dr ˜ n˜ j (q − r )πni (r − q)Φ(r ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ 1 ⎪ ⎭ Lkℓ,ni +n˜ j (s). ×1{n˜ j +ni ≥2} i! j!
(11.89)
11.2 Proof of Smoothness of (8.25) and (8.73)
259
Again, we consider first the case ni  = 0 and employ (11.70) with Z := Yλ (s). Integrating by parts in a5,n,2,Yλ (s) with πn˜ j (q − r ) 1j! Lkℓ,n˜ j (s), replacing Dˆ k,ℓ (s, q − p), we obtain in place of (11.72) ⎫ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ 0 0 ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ n n ⎪ × dq dq(∂ ˜ k,r ∂r G(u, r − q))(∂ℓ,r ∂r G(v, r − q))Y ˜ λ (s, q) ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ˜ n˜ j (q − r )Φ(r )dr ×Yλ (s, q)π ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ d ⎪ ∞ m/2
⎪ p( p − 1) p−2 µ −µu m/2−1 ⎪ ⎪ (Yµ,n,λ (s, r )) du e u =− ⎪ ⎪ m ⎪ 2 ( 2 − 1)! 0 ⎪ k,ℓ=1 ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ n × (∂k,r ∂r G(u, r − q))Yλ (s, q)πn˜ j (q − r )(∂ℓ,r Yµ,n,λ )(s, r )Φ(r )dr ⎪ ⎪ ⎪ ⎪ ⎪ ⎬ d p( p − 1)
µm − (Yµ,n,λ (s, r )) p−2 m 2 [( 2 − 1)!]2 ∞ ∞ k,ℓ=1 × du dv e−µ(u+v) u m/2−1 v m/2−1
=
∞ d µm/2 p
(Yµ,n,λ (s, r )) p−1 m du e−µu u m/2−1 2 − 1)! ( 0 2 k,ℓ=1 2 ∂rn G(u, r − q))Yλ (s, q)πn˜ j (q − r )Φ(r )dr × dq(∂kℓ,r
∞ d p
µm/2 du e−µu u m/2−1 (Yµ,n,λ (s, r )) p−1 m 2 ( − 1)! 0 2 k,ℓ=1 × dq(∂k,r ∂rn G(u, r − q))Yλ (s, q)(∂ℓ,r πn˜ j (q − r ))Φ(r )dr
+
d m/2 ∞ p
p−1 µ (Yµ,n,λ (s, r )) + du e−µu u m/2−1 m 2 ( − 1)! 0 2 k,ℓ=1 × dq(∂k,r ∂rn G(u, r − q))Yλ (s, q)πn˜ j (q − r )∂ℓ,r Φ(r )dr.
⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎭
(11.90) Let us now estimate the first term on the righthand side of (11.90). By the homogeneity of G(u, r − q) we have 2 2 ∂kℓ,r ∂rn G(u, r − q) = (−1)2+n ∂kℓ,q ∂qn G(u, r − q),
260
11 Proof of Smoothness, Integrability, and Itˆo’s Formula
whence, integrating by parts 2 ∂rn G(u, r − q))Yλ (s, q)πn˜ j (q − r ) dq(∂kℓ,r 2 = dq G(u, r − q)∂kℓ,q ∂qn (Yλ (s, q)πn˜ j (q − r ))
γn˜ (∂qn˜ Yλ (s, q))(∂qn+1k +1ℓ −n˜ πn˜ j (q − r )) = dq G(u, r − q) = +
˜ {n≤n+1 k +1ℓ }
dq G(u, r − q) dq G(u, r − q)
=: I + I I
˜ ˜ {n≤n+1 k +1ℓ ,n≤n}
˜ ˜ {n≤n+1 k +1ℓ ,n>n}
γn˜ (∂qn˜ Yλ (s, q))(∂qn+1k +1ℓ −n˜ πn˜ j (q − r )) γn˜ (∂qn˜ Yλ (s, q))(∂qn+1k +1ℓ −n˜ πn˜ j (q − r ))
In I the order of derivative at Yλ is ≤ n which need not be changed. In I I we simplify the notation in the terms to (∂qn˜ Yλ (s, q))G(u, r − q)∂qi πn˜ j (q − r ), ˜ In this case n ˜ equals either n + 1 or n + 2. In where i := n + 1k + 1ℓ − n. the first case, i = 1, in the latter case i = 0. We again integrate by parts once or twice against the product G(u, r − q)∂qi πn˜ j (q − r ) to reduce the order to n. The resulting expression in the product G(u, r − q)∂qi πn˜ j (q − r ) is a linear combination of terms of the following kind ˜j
ˆj
(∂q G(u, r − q))(∂q ∂qi πn˜ j (q − r )), ˜j = 2 − i − ˆj. Hence, the order in the derivative at G(u, r − q) is the same as the degree of the ˆj
resulting polynomial (∂q ∂qi πn˜ j (q − r )). Thus,
=
2 ∂rn G(u, r dq(∂kℓ,r
dq
− q))Yλ (s, q)πn˜ j (q − r )
⎫ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎬
γn, ˜ ˜j,ˆj ⎪ ⎪ ˜ ˜ ⎪ k +1ℓ ,n≤n} {˜j≤2−i−ˆj} {n+i=n+1 ⎪ ⎪ ⎪ ⎪ ˜j jˆ i n˜ ×(∂q G(u, r − q))(∂q ∂q πn˜ j (q − r ))(∂q Yλ (s, q)). ⎭
(11.91)
Employing Lemma 11.10,
∞
du e
0
≤ cβ,n,Φ,m
−µu m/2−1
2 dq(∂kℓ,r ∂rn G(u, r
u
c (where τ N (ω) = ∞, if sup Y (t)2 < ∞). 0≤t 0, Y0,N := Y0 1{Y0 0,Φ ≤N } , Y0,m,N := Y0,m 1{Y0,m 0,Φ ≤N }
(11.127)
11.3 Proof of the Itˆo formula (8.42)
271
implies that for p = 1, 2 and ∀N > 0:
2p 2p ⎫ sup EY (t, Y0,N ) − Y (t, Y0,m,N ) 0,Φ ≤ c p EY (t, Y0,N ) − Y (t, Y0,m,N ) 0,Φ ⎪ ⎪ ⎪ 0≤t≤T ⎪ ⎪ ⎪ ⎬ −→ 0 , as m −→ ∞ , 2p 2p sup EY (t, Y0,N )0,Φ ≤ c p EY (t, Y0,N )0,Φ , ⎪ ⎪ 0≤t≤T ⎪ ⎪ ⎪ 2p 2p ⎪ ⎭ sup EY (t, Y0,N )0,Φ ≤ c p EY (t, Y0,N )0,Φ , 0≤t≤T
(11.128)
where for the first inequality we used the bilinearity of (8.25). For Φ = ̟ , (11.128) follows from Theorem 8.6, since by the Cauchy–Schwarz inequality f 40,̟ ≤ c̟ f 40,4,̟ . For Φ ≡ 1, (11.128) follows from Kotelenez (1995b), Theorem 4.11. Note that for the truncated initial conditions the convergence assumption (11.126) also holds for the fourth moment of the norm of the difference of the initial conditions (by Lebesgue’s dominated convergence theorem). Since the Itˆo formula is a statement for almost all ω we may in what follows drop the subindex N and assume, without loss of generality, (11.128). Abbreviate Ym (t) := Ym (t, Y0,m ), Y (t) := Y (t, Y0 ). We know already from the first part (with smooth initial conditions) that the Itˆo formula holds for Ym (t)20,Φ . Denote the righthand side of (8.42) with process Ym (·) by Rm (t). If we replace on the righthand side of (8.42) all Ym terms by Y (t) the righthand side of (8.42) is a continuous semimartingale and will be denoted by R(t). We need to show that for all t ≥ 0, Y (t)20,Φ = R(t) with probability 1, which is precisely the Itˆo formula for Y (t)20,Φ . We easily see that the deterministic integrals in Rm (·) converge to the deterministic integrals of R(·) (in the topology of L 2,F ([s, T ]; W0,2,Φ )). Therefore, we will provide details only for the more difficult stochastic integrals. Abbreviate t Im,ℓ (t) := Ym2 (s, r )∂ℓ m ℓ (r, ds)Φ(r )dr 0
and Iℓ (t) :=
t
Y 2 (s, r )∂ℓ m ℓ (r, ds)Φ(r )dr.
0
The mutual quadratic variation
˜ r − r˜ )ds. d[∂ℓ m ℓ (r, ds), ∂ℓ m ℓ (˜r , ds)] = −(∂ℓ,r ℓ,r D)(s,
(11.129)
˜ r − r˜ ) in all variables (cf. (8.35)), we Thus, by the boundedness of (∂ℓ,r ℓ,r D)(s, obtain
272
11 Proof of Smoothness, Integrability, and Itˆo’s Formula 2
EIℓ (t) − Im,ℓ (t) ≤ c D˜ E
" t 0
(Y
2
(s, r ) − Ym2 (s, r ))Φ(r )dr
2 #
. (11.130)
A simple calculation shows that the righthand side of (11.130) can be estimated by c D,Φ ˜
t 0
(EY (s) − Ym (s)40,Φ )1/2 (EY (s) + Ym (s)40,Φ
1/2
.
Hence, by (11.128) in addition to (11.130) and the convergence assumption (11.126) (also holding after truncation for the fourth moment), 4 1/2 EIℓ (t) − Im,ℓ (t)2 ≤ c¯ D,Φ −→ 0 , as m −→ ∞ . ˜ (EY (0) − Ym (0)0,Φ ) (11.131)
Chapter 12
Proof of Uniqueness
Employing Itˆo’s formula, the basic estimate (8.48) is proved, which ensures strong uniqueness for quasilinear SPDEs under smoothness assumptions on the coefficients and initial conditions. We may without loss of generality assume s = 0 and that the coefficients are independent of t. Further, as in the smoothness proof of Chap. 11, the proof of (8.48) generalizes immediately to the coercive case, assuming in addition (8.74) and (8.75). Therefore, it suffices to provide the proof of (8.48) assuming σ ⊥ (r, µ, t) ≡ 0. Hence, the two solutions from Lemma 8.10 can be written as ⎫ t d ⎪ 1 t 2 ⎪ ⎪ Y (t) = Y0 + ∇ Y (s)F ·, Z (s) ds ⎪ ∂k,ℓ (Dk,ℓ (Z (s))Y (s))ds − ⎪ ⎪ 2 0 0 ⎪ ⎪ k,ℓ=1 ⎪ t ⎪ ⎪ ⎪ ⎪ − ∇(Y (s) J (·, p, Z (s))w(d p, ds)), ⎪ ⎬ 0
t d ⎪ ⎪ 1 t 2 ⎪ ˜ ˜ ˜ ˜ ⎪ ∂k,ℓ Dk,ℓ Z (s) Y (s) ds − ∇ Y˜ (s)F ·, Z˜ (s) ds ⎪ Y (t) = Y0 + ⎪ ⎪ 2 0 0 ⎪ ⎪ k,ℓ=1 ⎪ t ⎪ ⎪ ⎪ ⎪ ˜ ˜ ∇ Y (s) J ·, p, Z (s) w(d p, ds) . − ⎭ 0
(12.1)
Let c denote some finite constant whose values may change throughout the steps of a series of estimates. ∂ℓ denotes the partial derivative with respect to the ℓ’s spatial coordinate. If there are two spatial variables, we will write ∂ℓ,r to indicate that the partial derivative is to be taken with respect to the variable r , etc. Note that our assumptions imply by Theorem 8.6 that both Y and Y˜ are in L 0,F (C([0, ∞); W2,2,̟ )). Therefore, by Itˆo’s formula (8.42)
273
274
12 Proof of Uniqueness
⎫ ⎪ ⎪ ⎪ 0,̟ ⎪ ⎪ ⎪ ⎪ G F ⎪ t
d ⎪ ⎪ 2 ⎪ ˜ ˜ ˜ = ∂k,ℓ Dk,ℓ Z (s) Y (s) − Dk,ℓ Z (s) Y (s) , Y (s) − Y (s) ds ⎪ ⎪ ⎪ 0 k,ℓ=1 ⎪ ⎪ 0,̟ ⎪ ⎪ d ⎪ 2 t
⎪ ⎬ ∇ • Y (r, s)J (r, p, Z (s)) − Y˜ (r, s)J r, p, Z˜ (s) d p̟ (r )dr ds ⎪ +
< γ such that the imbedding H˜ −γ ⊂ H˜ −β is Hilbert Schmidt. On H˜ −β , W (·) is regular. We now extend (13.7) onto H˜ −β and solve it via (13.8) on H˜ −β . A typical example for A is the Laplacian on a bounded domain O ⊂ Rd such that its closure with respect to boundary conditions is a selfadjoint operator. (ii) Let A = 21 △ and O = Rd . We consider the Laplacian △ to be a closed operator on H := H0 = L 2 (Rd , dr ). Setting 2 (13.11) G(t, r ) := (2π t)−d/2 exp − r2t , the semigroup U (t) is given by
(U (t) f )(r ) =
G(t, r − q) f (q)dq,
(13.12)
where f ∈ H0 . In difference from the previous example, A and U (t) do not have a discrete spectrum.10 Therefore, a direct copy of the scale (13.10) does 9 10
Cf., e.g., Tanabe (1979), Sect. 2.3. The spectrum of 21 △ on Rd is the set (−∞, 0]. Cf., e.g., Akhiezer and Glasman (1950), Chap. VI.
13.1 Classification
297
not yield HilbertSchmidt imbeddings between Hilbert spaces on the scale. The remedy is found by using r 2 − △ instead of β − △, where r 2 is considered a multiplication operator. This operator has a discrete spectrum, and the CONS of eigenfunctions are the wellknown normalized Hermite functions. Thus, we define for γ ∈ R the fractional powers (r 2 − △)γ /2 . We now choose γ ≥ 0, denote corresponding Hilbert spaces by Hγ and obtain the Schwarz scale of distribution spaces11 S ⊂ Hγ ⊂ H0 ∼ = H′0 ⊂ H−γ ⊂ S ′ .
(13.13)
The imbedding ⊂ Hγ ⊂ Hβ is Hilbert Schmidt if, and only if, γ > β + d. Although U (t) does not commute with (r 2 − △)γ /2 it may still be extended and restricted to a strongly continuous semigroup on the corresponding spaces Hγ , which is enough to consider (13.7) on a corresponding Hilbert distribution space and solve it via (13.8). We have seen in both examples that linear SPDEs may be solvable for a large class of cylindrical Brownian motions by simply redefining them on a suitable Hilbert space of distributions. Langevin equations of type (13.7) arise naturally in central limit phenomena in scaling limits for the mass distribution of various particle systems. We refer the reader to Itˆo (1983, 1984) and to the papers by MartinL¨of (1976), Holley and Stroock (1978), Gorostiza (1983), Kotelenez (1986, 1988), Bojdecki and Gorostiza (1986), Gorostiza and Le´on (1990), and Dawson and Gorostiza (1990). There are many papers which deal with properties of infinite dimensional Ornstein–Uhlenbeck processes. Cf., e.g., Curtain (1981), Schmuland (1987), Bojdecki and Gorostiza (1991), Iscoe and McDonald (1989), and the references therein.
13.1.2 Bilinear SPDEs Let B(X, dW ) be a bilinear functional on H × H. The prototype of a bilinear SPDE can be described as follows: dX = AX dt + B(X, dW ), (13.14) X (0) = X 0 , Sometimes SPDEs of type (13.14) are also called linear in the literature. We find such a term confusing. We have already seen that for many selfadjoint operators A the linear SPDE may be redefined on a suitable distribution space where the original cylindrical Brownian motion becomes a regular one. In some cases of bilinear SPDEs, one may succeed in generalizing the variationsofconstants, considering the twoparameter random semigroup (also called random evolution operator) generated by the bilinear SPDE (cf. Curtain and Kotelenez (1987)). However, in 11
Cf. (15.32) in Sect. 15.1.3. The completeness of the normalized Hermite functions is derived in Proposition 15.8 of Sect. 15.1.3.
298
13 Comments on Other Approaches to SPDEs
many cases such a procedure may not be possible. Note that the bilinear SPDE with B(X, dW ) = ∇ X · dW is a special case of our equation (8.26) for J , Dkℓ independent of X and F ≡ 0, where the noise term can be cylindrical. For more examples of ¨ unel (13.14) with regular W (·), cf. Da Prato et al. (1982), Balakrishnan (1983), Ust¨ (1985), and the references therein. The difference between the linear SPDE (13.7) and the bilinear SPDE (13.14) is best explained by the following special case of (13.14): dX = AX dt + σ (X )dW ), (13.15) X (0) = X 0 , where σ is some nice function of x ∈ R and σ (X )dW is understood as a pointwise multiplication. We assume that W (t) is a standard cylindrical Brownian motion on H0 . Recall that, by our Fourier expansion (4.28), W (·) may be represented by our standard space–time white noise w(dr, dt).12 Further, to make things somewhat easier suppose that the semigroup generated by A, U (t) has a kernel G(t, r ) (the fundamental solution, associated with the Cauchy problem for A, I.e., U (t) is cast in the form of the following integral operator (U (t) f )(r ) := G(t, r − q) f (q)dq, where f ∈ H0 . In this case the mild solution of (13.15), if it exists, by definition, is the solution of t X (t) = G(t, r − q)X 0 (q)dq + G(t − s, r − q)σ (X (t, q)w(dq, ds). 0
(13.16) To see whether (13.16) can be well posed for nontrivial multiplication operators σ (·), we take a random field Y (t, q), measurable, adapted, etc. We compute the variance of the stochastic integral and obtain as a condition to solve the problem on H0 := L 2 (Rd , dr ) that the variance be finite, i.e., E =
t 0t 0
2 G(t − s, r − q)σ (Y (t, q)w(dq, ds) dr 2
(13.17)
2
G (t − s, r − q)σ (Y (s, q))dq ds dr < ∞.
For the important special case r 2 A = D△, and G(t, r ) := (4π Dt)−d/2 exp − 2Dt .
(13.18)
It is hard to guarantee a priori the condition (13.17). Take, e.g., σ (·) ≡ 1. We are then dealing with a special case of the linear SPDE (13.7) and the Integral on the righthand side of (13.17) equals ∞. Note that this problem would not have appeared 12
Cf. also (15.69) in Sect. 15.2.2.
13.1 Classification
299
if d = 1 and the SPDE would have been restricted to a bounded interval with Neumann or Dirichlet boundary conditions. Walsh (1986) shows that for d = 1 and G as in (13.15) the convolution integral G(t − s, r − q)w(dq, ds) is function valued. Following Kotelenez (1992a),13 we choose the weight function ̟ (r ) := (1 + r 2 )−γ with γ > d/2. This implies ̟ (r )dr < ∞. Instead of working with H0 = L 2 (R, dr ) we now analyze (13.16) on the weighted L 2 space H0,̟ = L 2 (R, ̟ (r )dr ). Young’s inequality14 implies that for d = 1 and G as in (13.18) the condition (13.17) holds, i.e., at least for bounded σ (·) the problem (13.16) may be well posed. The generalization to unbounded σ (·) and to d > 1 and higher order partial differential operators and pseudodifferential operators has been treated by Kotelenez (1992a). Finally, for the Laplacian on a bounded domain in Rd , closed with respect to Neumann boundary conditions, and d > 1 the linear convolution is necessarily distribution valued. In this problem G from (13.18) must be replaced by the corresponding fundamental solution.15 Finally, Nualart and Zakai (1989) show that (13.15) for A is in (13.18) and σ (x) ≡ x has a solution only in the space of generalized Brownian functionals if H0 := L 2 Rd , dr ), where d ≥ 2 and W (t) is the standard cylindrical Brownian motion on H0 . In other words, we cannot construct some state space (like S ′ ) and obtain solutions of (13.14) as ordinary random variables, defined on S ′ ).
13.1.3 Semilinear SPDEs Semilinear SPDEs may be written as dX = AX + B1 (·, X, ∇ X )dt + B(·, X, ∇ X )dW, X (0) = X 0 ,
(13.19)
The type (S L 0 , S L 0 , R) is probably the simplest and has been studied early on (cf. ChojnowskaMichalik (1976)). The difficulty increases if we consider (S L 1 , S L 0 , R) or (S L 1 , S L 1 , R). Both types have also been investigated over many years, and we wish to mention first of all the variational approach to those SPDEs in the thesis of Pardoux (1975) as well as Krylov and Rozovsky (1979), Gy¨ongy (1982), and the references therein. Chow and Jiang (1994) obtain space–time smooth solutions. For the semigroup approach see the paper by Da Prato (1982) and the book by Da Prato and Zabczyk (1992), which contains a rather complete description of those equations as well as many interesting references. 13 14 15
Cf. also our Chaps. 8, 9, and Sect. 15.1.4 for the use of H0,̟ . Cf. Theorem 15.7 in Sect. 15.1.2. Cf. Walsh (1986), Sect. 3.
300
13 Comments on Other Approaches to SPDEs
As previously mentioned in the comments on linear SPDEs, Dawson (1972) obtains existence, uniqueness and smoothness for a class of SPDEs of type (S L 0 , S L 0 , C), using a semigroup approach. The Brownian motion in Dawson’s paper is standard cylindrical, and his results hold for d = 1 if A is a secondorder elliptic partial differential operator. Dawson also considers higher dimensions and, accordingly, A is an elliptic partial differential operator of higher order. Marcus (1974) studies stationary solutions of the mild solution of an SPDE of type (S L 0 , S L 0 , C). Generalization of Dawson’s result have been obtained by Funaki (1983) for d = 1, Kotelenez (1987, 1992a,b) in a more general setting and by others. Kunita (1990) obtains solutions of SPDEs of type (S L 1 , S L 1 , C) and firstorder SPDE for regular and certain cylindrical Brownian motions which are equivalent to t our cylindrical Brownian motion W (t) := 0 J (·, q, s)w(dq, ds).16 Kunita’s approach is based on the method characteristics. Firstorder SPDEs are also analyzed by Gikhman and Mestechkina (1983), employing the method of characteristics. Krylov (1999) provides an analytic approach to solve a large class of semilinear SPDEs on L p spaces with p ≥ 2. For more work in this direction, cf. Kim (2005). We mention here Dawson’s (1975) paper in which he introduces the Dawson– Watanabe SPDE (to be commented on later). This SPDE describes a scaled system of branching Brownian motions. The driving term is standard cylindrical, A = △, and Dawson shows the existence of the process in higher dimensions as well, although the corresponding SPDE seems to be ill posed. For d = 1, however, one may solve the SPDE and obtain the Dawson–Watanabe process as a solution (cf. Konno and Shiga (1988) and Reimers (1989)). For W (·) being standard cylindrical, type (S L 1 , I, C) has a solution if d = 1, which can be shown by semigroup methods as well as variational methods (cf. Da Prato and Zabczyk (1992)). Note that the SPDEs driven by a Brownian motion and with deterministic coefficients are Markov processes in infinite dimensions with welldefined transition probabilities. Da Prato and Zabczyk (1991) analyze a semilinear SPDE of the type (S L 1 , I, R) or (S L 1 , I, C) so that our condition (13.9) holds. The semigroup, defined through the transition probabilities is called the Kolmogorov semigroup. This semigroup, P(t) is defined on Cb (H; R), the space of bounded continuous realvalued functions on H. Let f ∈ Cb (H; R) and δ and δ 2 denote the first and secondFr´echet derivatives with respect to the space variable X ∈ H. Then P(t) satisfies the following equation: ∂ ∂t (P(t) f )(X )
= 21 Tr (Q W δ2 (P(t) f )(X ) + &AX + B1 (X ), δ(P(t) f )(X )'H , (P(0) f )(X ) = f (X ).
(13.20)
Dropping initially the semilinear part, the SPDE becomes a linear SPDE. Its Kolmogorov semigroup, P2 (t), is shown to have certain smoothing properties (similar to analytic semigroups), and the Kolmogorov semigroup for the semilinear 16
Cf. (7.4) and (7.13) in Chap. 7.
13.1 Classification
301
SPDE may be obtained by variation of constants as the mild solution of a linear deterministic evolution equation: t (13.21) (P(t) f )(·) = (P2 (t) f )(·) + 0 P2 (t − s)(&B1 (·), δ(P(t) f )(·)'H ds.
For more details and results, cf. Da Prato (2004). For d = 2, Albeverio and R¨ockner (1991) employ Dirichlet form methods to weak solutions to an SPDE related to an SPDE in Euclidean field theory. Mikulevicius and Rozovsky (1999) analyze a class of SPDEs driven by cylindrical Brownian motions in the strong dual #′ of a nuclear space # (cf. our previous (13.10)). A large portion of the book by Kallianpur and Xiong (1995) treats SPDEs in a similar framework. Finally, let us also mention some qualitative results. Wang et al. (2005) consider two families of semilinear PDE’s. The first one is defined on a domain with (periodically placed) holes, and the second one is the weak (homogenization) limit of the first one, as the size of the holes tends to 0. Invariant manifolds for semilinear SPDEs with regular Brownian motion have been obtained by Duan et al. (2003, 2004).
13.1.4 Quasilinear SPDEs Quasilinear SPDEs can be written as dX = A(t, X )X + B1 (·, X, ∇ X )dt + B(·, X, ∇ X )dW, X (0) = X 0 ,
(13.22)
where for fixed f ∈ H A(t, f ) is an unbounded operator, e.g., an elliptic differential operator. Daletskii and Goncharuk (1994) employ analytical methods in the analysis of a special case of (13.22) with regular Brownian motion. Further, let us mention some results, obtained by a particle approach: Dawson and Vaillancourt (1995), Dawson et al. (2000), Kotelenez (1995a–c 1996, 1999, 2000), Kurtz and Protter (1996), Kurtz and Xiong (1999, 2001, 2004), Dorogovstev (2004b), and Dorogovtsev and Kotelenez (2006). Goncharuk and Kotelenez (1998) employ fractional steps in addition to particle and “traditional methods” to derive quasilinear SPDEs with creation and annihilation.
13.1.5 Nonlinear SPDEs For the type (NL,NL,R) we refer the reader to a paper by Lions and Souganidis (2000), who employ a viscosity solution approach. The noise in that setup consists of finitely many i.i.d. Rd valued Brownian motions. A direct generalization to the case of an infinite dimensional regular Brownian motion should not be too difficult.
302
13 Comments on Other Approaches to SPDEs
13.1.6 Stochastic Wave Equations Results and more references may be found in Carmona and Nualart (1988a,b), Marcus and Mizel (1991), Peszat and Zabczyk (2000), Dalang (1999), Dalang and Walsh (2002), Chow (2002), Dalang and Mueller (2003), Dalang and Nualart (2004), Dalang and SanzSol (2005), Dalang et al.(2006).
13.2 Models Applications of SPDEs within a number of different areas are reviewed. Historically, the development of SPDEs was motivated by two main models: • The Kushner17 and Zakai equations in nonlinear filtering (cf. Kushner (1967) and Zakai (1969)). • The Dawson–Watanabe equation for the mass distribution of branching Brownian motions (cf. Dawson (1975). Other models and equations followed. We will start with the nonlinear filtering equation.
13.2.1 Nonlinear Filtering Let r (t) be a Markov diffusion process in Rd which is described by the following SODE: t t r (t) = r0 + a(r (s), s)ds + b(r (s), s)dw(ds). ˜ 0
0
Suppose we observe the Rm valued process t h(r (s))ds + w(t). q(t) := 0
w(·) and w(·) ˜ are assumed to be independent standard Brownian motions with values in Rd and Rm , respectively. Under these conditions Kushner (1967) obtains a semilinear SPDE for the normalized conditional density of r (t) based on the observations of σ {q(s), s ≤ t}, where the linear secondorder operator A is the formal adjoint of the generator of r (·). Zakai (1969) obtains a bilinear SPDE for the unnormalized conditional density: 3 dX (t) = AY (t)dt + Q −1/2 h(·)X · q(dt), (13.23) X (0) = X 0 , where Q is the covariance operator of w(·) and X 0 is the density of the random variable r0 . 17
The Kushner equation is also known as the “Kushner–Stratonovich” equation.
13.2 Models
303
Besides the original papers by Kushner and Zakai we refer for more general nonlinear filtering equations to Pardoux (1979), Krylov and Rozovsky (1979), Da Prato (1982), Da Prato and Zabczyk (1992), Bhatt et al. (1995), and the references therein. Similar to the approach taken in this book, Kurtz and Xiong (2001) employ a particle method approach to the solution of the Zakai equation, and Crisan (2006) provides numerical solutions, based on particle methods.
13.2.2 SPDEs for Mass Distributions A number of papers have been devoted to stochastic reactiondiffusion phenomena. The SPDEs in many papers are semilinear of type (S L 0 , S L 0 , R), (S L 0 , S L 0 , C) as well as (S L 0 , I, R) and (S L 0 , I, C). However, the last two cases, if solvable, do not have nonnegative solutions and, therefore, cannot describe the distribution of some matter (cf. Kotelenez (1995b) and Goncharuk and Kotelenez (1998) for an alternative proof of the positivity of solutions). In contrast, a particle approach automatically has the positivity property. In addition, in a particle approach we describe the microscopic dynamics first before simplifying the time evolution of the mass distribution by passing to SPDEs and PDEs. There are at least two directions in the particle approach to SPDEs: 1. Branching Brownian motions and associated superprocesses 2. Brownian motions with meanfield interaction and associated McKean–Vlasov equations 1. We will first discuss superprocesses and the Dawson–Watanabe equation. Dawson (1975) analyzes an infinite system of i.i.d. diffusion approximations to a system of branching particles. For a set of atomic measures, he obtains the stochastic evolution equation on M f √ dX (t) = α X (t)dt + X (t)W (dt), (13.24) X (0) = X 0 . Here X 0 is supported by a countable set of points {ai } and α != 0 is the difference between birth and death rates. The square root is taken with respect to the mass √ at the point of support, W (t, ai ) are i.i.d. √ standard Brownian motions and X (t, ai ) is a factor at W (dt, ai ), whence X (t, ·) acts as a multiplication operator on W (dt, ·). The i.i.d assumption implies that W (t, ·) can be extended to a standard cylindrical Brownian motion (on H0 = L 2 (Rd , dr )). Through extension by continuity, Dawson extends the process from (13.24) to starts X 0 ∈ M f . The process (13.24) is a Markov process. Dawson then considers a pure diffusion, governed by the Laplacian △, which itself is a Markov process. The dynamics of both phenomena are linked through the Trotter product of their respective Markov semigroups, resulting in an M f valued Markov process with generator (A f ) = &(△ + α)δ f (X ), X ' + 21 Tr (δ2 f )(X ).
(13.25)
304
13 Comments on Other Approaches to SPDEs
Formally, this diffusion may be written as a solution to the following SPDE: √ dX (t) = (△ + α)X (t)dt + X (t)W (dt), (13.26) X (0) = X 0 .
There are two difficulties √ in the interpretation of (13.26). The first and most important problem is that X (t) is not defined for general measures. The second difficulty is that multiplication between measures and distributions is not defined. Recalling our analysis of the onedimensional case in (13.15)–(13.18) we conclude that (13.26) may be well posed for d = 1, and the only problem in (13.26) is posed by the square root, which is not Lipschitz. As mentioned earlier, Konno and Shiga (1988) and Reimers (1989) showed existence of a solution of (13.26) for d = 1. Consequently, the research activities on (13.26) “branched” into two directions: For d > 1 many papers have been written both on the generalization of (13.26) and on some qualitative properties. For d = 1, it became an active research area in SPDEs. In our earlier discussion of bilinear SPDEs we have seen that for d > 1 we cannot interpret (13.26) as an SPDE, and most of the work in that direction goes beyond the scope of this book. We just mention Perkins (1992, 1995) who includes interaction into the dynamics of branching Brownian particles. Further, Dawson and Hochberg (1979) and Iscoe (1988) analyze the support of measure valued branching Brownian motion. Cf. also Dawson (1993) for a detailed analysis, many properties and more references. Blount (1996) obtains for d = 1 a semilinear version of (13.25). For SPDEs of type (13.25), d = 1, cf. Mueller and Perkins (1992), DonatiMartin and Pardoux (1993), Shiga (1993), Mueller and Sowers (1995), Mueller and Tribe (1995), Mueller (2000), Dawson et al. (2003), and the references therein. 2. The second direction in particle methods is the central theme of this book. Apart from the author’s contributions (cf. Kotelenez (1995–2000)), we need to mention Vaillancourt (1988) and Dawson and Vaillancourt (1995) for a somewhat different approach. For generalizations of both the author’s and other results, cf. Wang (1995), Kurtz and Protter (1996), Goncharuk and Kotelenez (1998), Kurtz and Xiong (1999), Dawson et al. (2000), Kurtz and Xiong (2001), Dorogovtsev (2004b), and Dorogovtsev and Kotelenez (2006).
13.2.3 Fluctuation Limits for Particles We discussed earlier the infinite dimensional Langevin equations, which are obtained in central limit theorem phenomena for the mass concentration of particle systems. Kurtz and Xiong (2004) obtain an interesting generalization of this classical central limit theorem as follows: Consider the empirical process X N (·) associated with our system of SODEs (4.2) and the limit X (·) which is a solution of the quasilinear SPDE (8.54) in Chap. 8. Under some assumptions N1 (XN − X )(·) tends to the solution of a bilinear SPDE, whose coefficients depend on X (t). Giacomin
13.2 Models
305
et al. (1999) obtain a semilinear Cahn–Allen type SPDE with additive white noise for the fluctuations in a Glauber spin flip system. The space dimension is d = 1 for a rigorous result, and the equation is conjectured to be the limit also for d = 2. Solvability for d = 2 is not shown.
13.2.4 SPDEs in Genetics Fleming and Viot (1979) propose a derivation of the frequency distribution of infinitely many types in a population, following Dawson’s approach to measure valued diffusion. Donelly and Kurtz (1999) obtain a general particle representation which includes both some Fleming–Viot processes and the Dawson–Watanabe process. If the domain of types is the bounded interval [0, L] in the Fleming–Viot model, then this frequency distribution may be represented by the solution of the following SPDE: √ dX (t) = (△ + α)X (t) + g(X (t))dt + X (t)(1 − X (t)W (dt), (13.27) X (0, r ) = X 0 (r ) ∈ [0, 1], where the Laplacian √ is closed with respect to homogeneous Neumann boundary conditions. The term X (t)(1 − X (t)W (dt) is called the random genetic drift. For W (·) regular (13.27) was solved by Viot (1975) (even in higher dimensions). For W (t) being standard cylindrical in L 2 ([0, L]; dr ), this equation is of the type we discussed in Sect. 13.2.2.
13.2.5 SPDEs in Neuroscience Walsh (1981) represents a neuron by the onedimensional interval [0, L]. He obtains an equation for an electrical potential, X (t, r ), by perturbing the (linear) cable equation by impulses of a current. These impulses are modeled by space–time white noise times a nonlinear function of the potential: ⎫ dX (t, r ) = (△ − 1)X (t, r ) + g(X (t, r ), t)w(dr, dt), ⎪ ⎪ ⎬ ∂ ∂ (13.28) ∂r X (0, t) = ∂r X (L , t) = 0, ⎪ ⎪ ⎭ X (0, r ) = X 0 (r ).
Under a Lipschitz assumption on g, Walsh obtains existence and uniqueness for the mild solution of (13.28) as well as space–time smoothness and a version of a multiparameter Markov property. It was in part this work which motivated Walsh’s general approach to semilinear SPDEs, driven by space–time white noise and its generalization, employing variation of constants (cf. Walsh, 1986). Kallianpur and Wolpert (1984) analyze an Ornstein–Uhlenbeck approximation to a similar stochastic model, where the neuron can be represented by a compact ddimensional manifold and the driving noise is a generalized Poisson process.
306
13 Comments on Other Approaches to SPDEs
13.2.6 SPDEs in Euclidean Field Theory These are semilinear SPDEs driven by space–time white noise, i.e., they are of type (S L 1 , I, C). As we have already seen in our comments, these equations can be solvable in space dimension d = 1 by direct SPDEs methods. The nonlinear term is usually a polynomial g(X ) and in higher dimension the powers X m (r ) have to be replaced by the Wick powers: X m (r ):18 The resulting quantization equation is written as ⎫ dX (t) = (△X (t)+ : g(X (t)) :)dt + W (dt), ⎪ ⎪ ⎬ X (0) = X 0 , (13.29) ⎪ ⎪ ⎭ and possible selfadjoint boundary conditions. For existence, uniqueness, and properties in space dimension d = 2 we refer to papers of Albeverio and R¨ockner (1991), Albeverio et al. (2001), Da Prato and Debussche (2003), and the references therein. A version of (13.29) with a regular H0 valued Brownian motion and with an ordinary polynomial has been analyzed in detail by Doering (1987). This paper also contains a number of interesting comments on the problem of stochastic quantization and Euclidean field theory.
13.2.7 SPDEs in Fluid Mechanics Monin and Yaglom (1965) treat the statistical approach to turbulence. The velocity field of a fluid is described by a space–time vectorvalued random field. If the random field is stationary, homogeneous, and isotropic, the correlation matrix has a particularly simple structure which can be compared with empirical data (cf. Monin and Yaglom (loc. cit. Chap. 7). Observe that the velocity field is governed by the Navier–Stokes equations. Solutions of a forced version of these equations with random initial conditions may yield stationary, homogeneous and isotropic random fields as solutions. This begs the question whether the solutions of a suitably forced version of the Navier–Stokes are consistent with theoretical considerations, based on the theory of Kolmogorov and Obukhov and its generalizations and whether their correlation matrices are structurally similar to the empirical data. If the fluid is incompressible and the forcing is a Gaussian random field which is white in time, we obtain stochastic Navier–Stokes equations. We recommend the review paper by Kupiainen (2004) and the references therein as a starting point for more information on turbulence, homogeneous, and isotropic solutions and (in 2D) on ergodicity results of stochastically forced Navier–Stokes equations. Let U (t, r ) be the velocity of the fluid. The stochastic Navier–Stokes equations, mentioned above, are cast in the form
18
Cf. Da Prato and Zabczyk (1992), Introduction, 06, for a definition.
13.2 Models
307
⎫ dU (t) = (ν△U (t) − (U · ∇)U − ∇ p)dt + W (dt), ⎪ ⎪ ⎪ ⎪ ⎪ ⎬ ∇ · U = 0, U (0) = U0 ,
and possible boundary conditions.
⎪ ⎪ ⎪ ⎪ ⎪ ⎭
(13.30)
Here, p is the pressure of the fluid and ν the kinematic viscosity. A first step in the derivation of stationary, homogeneous, and isotropic random field solutions of (13.30) is to show existence of a stationary solution of (13.30) and, if possible, ergodicity. Stochastic Navier–Stokes equations of type (13.30) were considered by Chow (1978) for regular (vectorvalued) Brownian motion. Vishik and Fursikov (1980) show existence and uniqueness (for d = 2) and derive a number of properties. Flandoli and Maslowski (1995) obtain an ergodicity result for a 2D version of (13.30). For more results in the 2D case, cf. Mattingly and Sinai (2001) as well as Da Prato and Debussche (2002) and the references therein. In the paper by Da Prato and Debussche the Brownian motion is standard cylindrical. Da Prato and Debussche (2003) show existence of a solution and ergodicity for d = 3, where the Brownian motion is regular. Mikulevicius and Rozovsky (2004) consider the fluid dynamics of a fluid particle as a stochastic flow. The time derivative of the flow is governed by the velocity field of a more general version of (13.30) and by a stochastic Stratonovich differential. The last term represents the fast oscillating component of the flow and shall model turbulence. Further, existence and uniqueness results with regard to the stochastic Navier–Stokes equations with statedependent noise are obtained by Mikulevicius and Rozovsky (2005). For convergence of the solutions of the NavierStokes equations toward random attractors, we refer the reader to Schmalfuss (1991), Crauel et al.(1997), Brze´zniak and Li (2006), and the references therein. Kotelenez (1995a) analyzes a stochastic Navier–Stokes equation for the vorticity of a 2D fluid,19 employing a particle approach. Apart from the fact that, in the approximation, the (positive) point masses are replaced positive and negative intensities (for the angular velocities), this equation is a semi–linear version of our SPDE (8.26). Sritharan and Sundar (2006) consider a 2D Navier–Stokes equations perturbed by multiplicative noise and establish a large deviation principle for the small noise limit. Neate and Truman (2006) employ the ddimensional stochastic Burgers equation in the study of turbulence and intermittence (cf. also the references therein). Baxendale and Rozovsky (1993) consider an SPDE to study the longtime asymptotics of a magnetic field in a turbulent δcorrelated flow of an ideal incompressible fluid.
19
Cf. Chap. 8, (8.15).
308
13 Comments on Other Approaches to SPDEs
13.2.8 SPDEs in Surface Physics/Chemistry In many formulations, SPDEs arising from the study of interfaces between two phases (e.g., vapor–liquid interface) are quasilinear, where the quasilinearity results from the motion by mean curvature of the interface. Such an SPDE for a sharp interface has been analyzed by Yip (1998), where the stochastic perturbation is given by Kunitas Gaussian martingales increments dM(t, r ). This is equivalent to our choice of G(r, q, t)w(dq, dt), as shown in Chap. 7. Yip obtains existence of a solution using fractional steps. Souganidis and Yip (2004) show that the solution of a similar SPDE can be unique, where the deterministic motion may have several solutions. Funaki and Spohn (1997) derive a sharp interface model from an interacting and diffusing particle model (on a lattice) as a macroscopic limit, employing a hydrodynamic scaling. Replacing the i.i.d. Brownian motions (indexed by the lattice sites) in the paper by Funaki and Spohn by correlated Brownian motions should result in a sharp stochastic interface, which itself should be the solution of a quasilinear SPDE, similar to the SPDE obtained by Yip (loc. Cit.). Bl¨omker (2000) analyzes surface growth and phase separation, whose macroscopic regime is often described by the semilinear fourthorder parabolic Cahn– Hilliard equation.
13.2.9 SPDEs for Strings For SPDEs describing the motion of strings we refer to Funaki (1983), Faris and JonaLasinio (1982), Mueller and Tribe (2002), and the references therein.
13.3 Books on SPDEs To our knowledge Rozovsky’s book in 1983 is first book on SPDEs. It treats bilinear SPDEs driven by regular Brownian motion with emphasis on filtering. Metivier (1988) obtains weak solutions for a class of semilinear SPDEs. Da Prato and Zabczyk (1992) develop linear, bilinear, and semilinear SPDEs using a semigroup approach. Both regular and cylindrical Brownian motions appear as stochastic driving terms. It also contains a chapter on invariant measures and large deviations as well appendixes on linear deterministic systems and control theory. The problem of invariant measures and ergodicity is treated in detail by Da Prato and Zabczyk (1996). Kallianpur and Xiong (1995) analyze linear, bilinear and some semilinear SPDEs driven by regular and cylindrical Brownian motions and quasilinear SPDEs driven by Poisson random measures. This book also provides a chapter on large deviations. Greksch and Tudor (1995) analyze semilinear SPDEs, employing both a semigroup approach and variational methods. Flandoli (1995) obtains regularity results for stochastic flows associated with the solutions of bilinear SPDEs. The book
13.3 Books on SPDEs
309
by Holden et al. (1996) deviates from all the previously mentioned monographs in that the state space is a space of generalized random variables (cf. the paper by Nualart and Zakai (1989)) and multiplication is replaced by the Wick product (cf. (13.29) above and the references quoted). Peszat and Zabczyk (2006) employ a semigroup approach to SPDEs, driven by L´evy processes with applications to models in physics and mathematical finance. Knoche–Prevot and R¨ockner (2006) provide an introduction to the variational approach. Finally, we must mention Walsh (1986) and Dawson (1993). Although those St. Fleur notes have not appeared as separate books, they are as comprehensive and detailed as any good monograph. Walsh employs a semigroup approach, but uses a space–time formulation for the stochastic terms instead of infinite dimensional Brownian motions or martingales. We have adopted the integration with respect to space–time white noise from Walsh. Dawson’s St. Fleur notes are, strictly speaking, not on SPDEs. However, the particle approach to SPDEs has been strongly influenced by Dawson’s work.
Chapter 14
Partial Differential Equations as a Macroscopic Limit
14.1 Limiting Equations and Hypotheses We define the limiting equations and state additional hypotheses. In this section we add the subscript ε to the coefficients of the SODEs and SPDEs which are driven by correlated Brownian noise. Further, we restrict the measures to M1 , since we want to use results proved by Oelschl¨ager (1984) and G¨artner (1988) on the derivation of the macroscopic McKean–Vlassov equation as the limit of interacting and diffusion particles driven by independent Brownian motions. To this end, choose a countable set of i.i.d. Rd valued standard Brownian motions {β i } j∈N . Let F0 (r, µ, t) and J0 (r, µ, t) be Rd valued and Md×d valued functions, respectively, jointly measurable in all arguments. Further, suppose (rℓ , µℓ , t) ∈ Rd × M × R, ℓ = 1, 2, Let c F , cJ ∈ (0, ∞) and assume global Lipschitz and boundedness conditions ⎫ F0 (r1 , µ1 , t) − F0 (r2 , µ2 , t) ≤ c F {ρ(r1 − r2 ) + γ (µ1 − µ2 )} ⎪ ⎪ ⎪ d ⎪
⎪ ⎪ 2 ⎪ ⎪ (J0,kℓ (r1 , µ1 , t) − J0,kℓ (r2 , µ2 , t)) ⎪ ⎪ ⎬ k,ℓ=1 ! (14.1) 2 ρ 2 (r1 − r2 ) + γ 2 (µ1 − µ2 ) , ≤ cJ ⎪ ⎪ ⎪ ⎪ ⎪ d ⎪
⎪ 2 2 ⎪ ⎪ J0,kℓ (r, µ, t) ≤ c F,J . F0 (r, µ, t) + ⎪ ⎭ k,ℓ=1
Consider stochastic ordinary differential equations (SODEs) for the displacement of r i of the following type ⎫ i i i (t) = F0 (r0,N (t), X N (t), t)dt + J0 (r0,N (t), X N (t), t)dβi (t), ⎪ dr0,N ⎬ N
(14.2) ⎪ r i (s) = q i , i = 1, . . . , N , X 0,N (t) := m i δr i (t) . ⎭ 0,N i=1
The subscript 0 indicates that there is no correlation between the Brownian noises for each particles. The Lipschitz conditions are those of Oelschl¨ager. G¨artner allows 313
314
14 Partial Differential Equations as a Macroscopic Limit
for local Lipschitz conditions. Both authors restrict the analysis to timeindependent coefficients. However, assuming conditions, uniformly in t, we may assume that their results can be formulated also for the nonautonomous case. The twoparticle diffusion matrix of the noise is given by i = j, J0 (r i , µ, t)J0T (r j , µ, t), if (14.3) D˜ 0 (µ, r i , r j , t) := 0, if i != j, where AT is the transpose of a matrix A. Set D0 (µ, r, t) := D˜ 0 (µ, r, r, t).
(14.4)
Under the above assumptions X 0,N (·) ⇒ X 0,∞ (·) in C([0, ∞); M1 ), as n −→ ∞,
(14.5)
where X 0,∞ (·) is the unique solution of the macroscopic McKean–Vlassov equation (or “nonlinear diffusion equation”):1 ⎫ d ⎪ ∂ 1 2 ⎬ X 0,∞ = ∂kℓ (D0,kℓ (X 0,∞ , ·, t)X 0,∞ ) − ▽ · (X 0,∞ F0 (·, X 0,∞ , t)), ∂t 2 k,ℓ=1 ⎪ ⎭ X 0,∞ (0) = µ. (14.6)
Along with (14.6) we define infinitely many SODEs, whose empirical distribution is the solution of (14.6): ⎫ i i i dr0,∞ = F0 (r0,∞ , X 0,∞ , t)dt + J0 (r0,∞ , X 0,∞ , t)dβ i , ⎪ ⎬ N
1 (14.7) i i r0,∞ (0) = q , i = 1, 2, . . . , X 0,∞ (t) = lim δ i (t). ⎪ ⎭ N →∞ N r0,∞ i=1
Existence and uniqueness for (14.7) follow under the assumptions (14.1).2 Let σε (r, µ, t) be the nonnegative square root of Dε (µ, r, t). Since the entries of Dε (µ, r, t) are bounded, uniformly in (r, µ, t), the same boundedness holds for σε (r, µ, t).3
Hypothesis 14.1 (i) Suppose that σi⊥ ≡ 0 ∀i in (4.10). (ii) Suppose that for each r ∈ Rd , t ≥ 0 and µ ∈ M1 , σε (r, µ, t) is invertible. 1 2 3
Cf. Oelschl¨ager (1984). G¨artner (loc. cit.) obtains the same result under local Lipschitz and linear growth conditions, making an additional assumption on the initial condition. Cf. Oelschl¨ager (1984), G¨artner (loc. cit.), and Kurtz and Protter (1996). Cf. also Remark 14.1. In Sect. 4.3, we provide a class of coefficients which satisfy Hypothesis 14.1.
14.1 Limiting Equations and Hypotheses
315
(iii) Further, suppose that for any compact subset K of Rd , any compact subset C ⊂ M1 , any T > 0 and any δ > 0 the following three relations hold: lim sup sup sup (Fε (r, µ, t) − F0 (r, µ, t) + σε (r, µ, t) − J0 (r, µ, t)) = 0, ε↓0 r ∈K 0≤t≤T µ∈C
(14.8)
where  ·  denotes the Euclidean norms in Rd and Md×d , respectively. (iv) sup
σε−1 (r, µ, t) < ∞,
(14.9)
 D˜ ε (µ, r, q, t) = 0, 4
(14.10)
sup
1≥ε>0 r ∈K,0≤t≤T,µ∈C
(v) lim sup
sup
ε↓0 r −q>δ 0≤t≤T,µ∈C
(vi) lim sup sup
sup
η↓0 r ∈K µ∈C 0≤s,t≤T,t−s≤η
F0 (r, µ, t) − F0 (r, µ, s) + J0 (r, µ, t)
−J0 (r, µ, s) = 0,
(14.11)
i.e., F0 (r, µ, t) and J0 (r, µ, t) are continuous in t, uniformly in (r, µ) from compact sets K × C. (vii) Given the solution X 0,∞ (t) of (14.7), suppose that D0,kℓ (X 0,∞ (·), ·, ·) and F0,k (·, X 0,∞ (·), ·) as functions of (r, t) are both twice continuously differentiable, where k, ℓ = 1, . . . , d. ⊔ ⊓ Remark 14.1. (i) In what follows we assume the Oelschl¨ager–G¨artner result to hold also for nonautonomous coefficients, since one can obviously generalize their results, assuming that their hypotheses hold uniformly in t. (ii) A closed kdimensional submanifold of Rn will be called nonattainable by an Rn valued diffusion, (r1 (·, q1 ), . . . , rn (·, qn )) with initial values (q1 , . . . , qn ) if P {(q1 , . . . , qn ) ∈ } = 0 implies P {∪t>0 (r1 (·, q1 ), . . . , rn (·, qn )) ∈ )} = 0. (14.12) ⊔ ⊓ Friedman (1976), Sect. 11, obtains general conditions for nonattainability of closed kdimensional submanifold of Rn for linear Rn valued diffusions, under the assumption that the coefficients are autonomous and twice continuously differentiable. In Lemma 14.8, we extend Friedman’s result to the nonautonomous case, defined by the first m diffusions of the system (14.7). The standard device of adding the time t as another dimension converts the nonautonomous system into an autonomous (but degenerate) system. 4
Cf. the definition of D˜ ε (µ, r, q, t) before (8.23).
316
14 Partial Differential Equations as a Macroscopic Limit
The differentiability assumption (iv) on the coefficients as functions of (r, t) in Hypothesis 14.1 depends on the existence of weak derivatives of the solution of the McKean–Vlasov equation (14.7) with respect to t. For an example where the assumption holds, cf. (14.3) (v). Finally, let X ε,∞ (0) be the continuum limit of X ε,N (0) := Z 0,∞ (Y, t, X ε,∞ (0)) be the solution of the random media PDE:
N
i=1
m i δqεi and ⎫
d ⎪ ∂ 1 2 ⎬ Z 0,∞ (Y ) = ∂kℓ (D0,kℓ (Y, ·, t)Z 0,∞ (Y )) − ▽ · (Z 0,∞ (Y )F0 (·, Z 0,∞ (Y ), t)), ∂t 2 k,ℓ=1
Z 0,∞ (0) = µ,
⎪ ⎭
(14.13)
where Y (·) ∈ Mloc,2,0,∞ .
14.2 The Macroscopic Limit for d ≥ 2 Assuming d ≥ 2 we prove that, under the hypotheses from 14.1 and Chap. 4, the solutions of the SPDE (8.26) converge weakly toward the solution of a corresponding quasilinear macroscopic PDE, as the correlation length tends to 0. For m ∈ N, set
! m := ( p 1 , . . . , p m ) ∈ Rd·m : ∃i != j, i, j ∈ {1, . . . , m}, with pi = p j .
Infinite sequences in Rd will be denoted either (r 1 , r 2 , . . .) or r (·) . The corresponding state space, (Rd )∞ , will be endowed with the metric d∞ (r (·) , q (·) ) :=
∞
k=1
2−k ρ(r k − q k ).
Theorem 14.2. 5 Suppose Hypotheses 4.1 and 14.1. Further, suppose that d ≥ 2 and that {qε1 , qε2 , . . .} is a sequence of exchangeable6 initial conditions in (4.10) and (14.2), respectively, such that for all m ∈ N and ε ≥ 0, 5 4 P (qε1 , . . . , qεm ) ∈ m = 0, where qε1 , . . . , qεm are the initial conditions in (4.10) for ε > 0 and in (14.2) for ε = 0, respectively. Finally, suppose 5 6
This theorem was obtained by Kotelenez and Kurtz (2006), and, in what follows, we adopt the proof from that paper. By definition exchangeability means that {qε1 , qε2 , . . .} ∼ {qεπ(1) , qεπ(1) , . . .} whenever π is a finite permutation of N. Cf. Aldous (1985).
14.2 The Macroscopic Limit for d ≥ 2
317
(X ε (0), qε1 , qε2 , . . .) ⇒ (X 0,∞ (0), q01 , q02 , . . .) in M1 × (Rd )∞ , as ε ↓ 0 , N 1 δq j N N →∞ j=1 0
where X ε (0) are the initial values for (8.26) and X 0,∞ (0) = lim initial condition of (14.6). Then
is the
X ε ⇒ X 0,∞ in C([0, ∞); M1 ), as ε ↓ 0
(14.14)
Z ε (Y ) ⇒ Z 0,∞ (Y ) in C([0, ∞); M1 ), as ε ↓ 0 ,
(14.15)
and where Z ε (Y ) is the solution of the SPDE (8.25) with coefficients Fε and Dε and Y (·) as in (14.13). The proof of Theorem 14.2 will be the consequence of a series of lemmas, which we will derive first. 2 Lemma 14.3. Let (qε1 , . . . , qεm , . . .) be a finite or infinite subsequence of initial conditions for (4.10) at time 0 and X ε (0) an initial condition for (8.26) such that (X ε (0), qε1 , qε2 , . . .) satisfy the assumptions of Theorem 14.2. Then (X ε (·), rε (·, X ε , qε1 ), . . . , rε (·, X ε , qεm , . . .)) is relatively compact in C([0, ∞); M1 × (Rd )∞ ). Proof. (i) It sufficient to restrict the sequences to the first m coordinates in the formulation of the Lemma. (ii) For s ≤ t ≤ T , E(γ 2 (X ε (t), X ε (s))/Fs ) 2 = E sup ( f (r (t, X ε , s, q)) − f (q))X ε (s, dq) /Fs ) f L ,∞ ≤1
(by (15.22) and (8.50)) ≤ E(
ρ(rε (t, q), q)X ε (s, dq))2 /Fs )
(since the f , used in the norm γ , are Lipschitz with Lipschitz constant 1) ≤ E( ρ 2 (rε (t, q), q)X ε (s, dq)/Fs ) (by the Cauchy–Schwarz inequality and X ε (s, Rd ) = 1) = E(ρ 2 (rε (t, q), q)/Fs )X ε (s, dq) ≤ cT (t − s)X ε (0, Rd ) = cT (t − s)
independent of ε by the boundedness of Fε and Jε (cf. (4.11)) and by X ε (s, Rd ) = X ε (0, Rd ) = 1 a.s. Similarly, for the mparticle process with a constant cT,m .
318
14 Partial Differential Equations as a Macroscopic Limit
(iii) To prove relative compactness of the marginals we must show tightness. Indeed, for the mparticle process this is obvious and its tightness follows from (4.11). For the (M1 , γ )valued process this follows from the proof of a similar statement, when M1 is endowed with the Prohorov metric and the equivalence of γ and the Prohorov metric.7 To show tightness of X ε (t) let δ > 0 be given. Choose a closed ball SL in Rd with center 0 and radius L = L(T ) such that for all ε > 0 4 5 (α) P X ε (0, SLc ) > 2δ ≤ 2δ , & 2 T −2 (β) E sup L Fε (rε¯ (s, q), X ε¯ (s), s)ds 0 q∈Rd ,ε>0 ⎫ d T ⎬ δ
2 + (rε (s, q), p, X ε (s), s)d p ds ≤ . Jε,kℓ ⎭ 2 0 k,ℓ=1
Cf. (4.11)–here we used the abbreviation Ac := Rd \A for A ∈ B d . Then, 3 c (r ε (t, q))X ε (0, dq) > δ P 1 S2L 3 ≤P (1 SLc (rε (t, q) − q) + 1 SLc (q))X ε (0, dq) > δ 3 δ rε (t, q) − q2 F X (0, dq) + ≤E E ε 0 2 L2 < 2δ + 2δ = δ.
Employing Theorems 15.26 and 15.27 of Sect. 15.2.1 completes the proof. ⊓ ⊔
In what follows let (X ε , rε1 , . . . , rεNε ) := (X ε,Nε , rε (·, X ε,Nε , q 1 ), . . . , rε (·, X ε,Nε , q Nε )) be a convergent subsequence in CM1 ×(Rd )∞ [0, ∞), where ε ↓ 0. We allow Nε = ∞. If Nε < ∞, we require Nε → ∞. Set t σε−1 (rεi (s), X ε (s), s)Jε (rεi (s), q, X ε (s), s)w(dq, ds), (14.16) wεi (t) := 0
where X ε := X ε,Nε is the empirical process for (rε1 , . . . , rεNε ). By Levy’s theorem8 each wεi is an Rd valued Brownian motion. Abbreviating cε (r, q, µ, t) := σε−1 (r, µ, t) D˜ ε (µ, r, q, t)σε−1 (q, µ, t), we obtain [wεi , wεj ](t) =
0
t
cε (rεi (s), rεj (s), X ε (s), s)ds,
(14.17)
where the lefthand side of (14.17) is the tensor quadratic variation.9 Moreover, hypotheses (14.9) and (14.10) imply for any δ > 0, any compact subset C ⊂ M1 and any T 7 8 9
Cf. Sect. 15.2.1, Theorems 15.23 and 15.24. Cf. Theorem 15.37, Sect. 15.2.3. Cf. Sect. 15.2.3.
14.2 The Macroscopic Limit for d ≥ 2
lim sup
319
sup
ε↓0 r −q>δ µ∈C,0≤t≤T
cε (r, q, µ, t) = 0.
(14.18)
Define for 1 > δ > 0 and m < Nε stopping times τεm (δ) :=
min
1≤i!= j≤m
inf{t ≥ 0 :
rεi (t) − rεj (t) ≤ δ}.
(14.19)
Let G be the σ algebra, generated by w(dr, dt) and by all events from F0 , i.e., we set ∞ 3 G := σ 1 B (r, t)w(dr, dt), B ∈ B d+1 , A ∈ F0 , Rd
0
B d+1
where is the Borel σ algebra in Rd+1 . Further, let {β i }i∈N be i.i.d. Rd valued Brownian motions, defined on (Ω, F, Ft , P) and independent of G. Set
t Nε N 1 βε,δ,m (t) := (wε1 , . . . , wε ε )(t ∧ τεm (δ))+ 1{u≥τεm (δ)} (dβ 1 , . . . , dβ Nε )(u). , . . . , βε,δ,m 0
(14.20) Nε 1 Clearly, (βε,δ,m , . . . , βε,δ,m ) is a continuous square integrable Rd·Nε valued martingale with tensor quadratic variation t / 1 / 1 0 Nε 0 w , . . . , wεNε (t ∧ τ m (δ)) + β (t) = 1{u≥τεm (δ)} du Id·Nε . , . . . , β ε ε ε,δ,m ε,δ,m 0
(14.21) Here Ik denotes the identity matrix on Rk and for Nε = ∞ Id·∞ is the identity operator in (Rd )∞ . Lemma 14.4. For each δ > 0 and m < Nε 1 m (βε,δ,m , . . . , βε,δ,m ) ⇒ (β 1 , . . . , β m ) in C([0, ∞); Rm·d ), as ε ↓ 0.
(14.22)
Proof. The family of marginals {X ε (t), ε > 0}, t ≥ 0, is relatively compact by Lemma 14.3. So, by the compact containment condition10 in addition to Lebesgue’s dominated convergence theorem / 1 0 m β ε,δ,m , . . . , βε,δ,m (t) → t Id·m , as ε ↓ 0 in probability. Hence, (14.22) follows from the martingale central limit theorem.11
⊔ ⊓
Consider the following system of SODEs ⎫ i i i i drε,δ,m,N = Fε (rε,δ,m,N , X ε )dt + σε (rε,δ,m,N , X ε )dβε,δ,m ⎪ ε ε ε ⎬ Nε
1 i i rε,δ,m,Nε (0) = qε , i = 1, . . . , Nε , X ε,δ,m,Nε (t) := δi (t), ⎪ ⎭ Nε rε,δ,m,Nε i=1 (14.23) 10 11
Cf. Remark 7.3 in Sect. 3 of Ethier and Kurtz (1986) and (14.24). Cf. Theorem 15.39, Sect. 15.2.3.
320
14 Partial Differential Equations as a Macroscopic Limit
where for Nε = ∞ here and in what follows i = 1, . . . , Nε shall mean i ∈ N. 1 (·), . . .) Repeating the proof of Lemma 14.3, we show that (X ε,δ,m,Nε (·), rε,δ,m,N ε is relatively compact in C([0, ∞); M1 × R∞ ) and, therefore, we may, without 1 loss of generality, assume that (X ε,δ,m,Nε (·), rε,δ,m,N (·), . . .) converges weakly. Let ε 1 (r0,∞ , . . .) be the solution of (14.7). Lemma 14.5. For every δ > 0, m < Nε Nε 1 1 ) ⇒ (r0,∞ (rε,δ,m,N , . . . , rε,δ,m,N , . . .) in C([0, ∞) : (Rd )∞ ), as ε ↓ 0 . ε ε
Proof. Recall that (qε1 , . . . , qεNε ) ⇒ (q01 , . . .) in (Rd )∞ , as ε ↓ 0. Using the boundedness of the coefficients in (14.23) we may repeat the proof of Lemma 14.3 and conclude that the family X ε,δ,m,Nε (·) is relatively compact in C([0, ∞); M1 ), as a function of ε, δ, and m. Hence, the aforementioned compact containment condition of Ethier and Kurtz (loc. cit.) now reads: For every L ∈ N there is a compact subset C L of M1 such that inf
{ε>0,δ>0,m 0, δ > 0, m ∈ N with probability 1 ∪0≤t≤T ({X ε (t, ω)} ∪ {X ε,δ,m,Nε (t, ω)} ∪ {X 0,∞ (t, ω)) ⊂ ∪ L∈N C L .
(14.25)
For each N ∈ N define maps Ψ N from ((Rd )∞ , d∞ ) into (M1 , γ f ) by N 1
Ψ N r (·) = δr i , N i=1
where here and in what follows (r 1 , . . . , r N ) are the first N coordinates of r (·) . The definition of the norm γ implies that for two sequences r (·) and q (·) N 1
̺(r i − q i ). γ (Ψ N r (·) − Ψ N q (·) ≤ N
(14.26)
i=1
So the maps Ψ N are continuous. Further, set
D∞ := {r (·) ∈ (Rd )∞ : such that lim Ψ N (r (·) ) exists}, N →∞ & (·) ), if lim Ψ (r r (·) ∈ D∞ , N Ψ∞ (r (·) ) := N →∞ δ0 , if r (·) ∈ / D∞ . This definition implies that D∞ is a Borel set in ((Rd )∞ , d∞ ) and that Ψ∞ is a measurable map from ((Rd )∞ , d∞ ) into (M1 , γ f ). Further, by the previously menN tioned density of measures of the form N1 δr i in (M1 , γ ) i=1
14.2 The Macroscopic Limit for d ≥ 2
321
Ψ∞ (D∞ ) = M1 . Finally, we set 4 5 D∞,L := r (·) ∈ D∞ : Ψ∞ (r (·) ) ∈ C L .
Define stopping times
4 5 τε (δ, m, L) := inf t ≥ 0 : X ε,δ,m,Nε (t) ∈ / CL .
(14.27)
By (14.25), ∀ε > 0, δ > 0, m ∈ N with probability 1:
τε (δ, m, L) ↑ ∞, as L −→ ∞.
(14.28)
If X ε (0) ∈ C L with probability 1 for some L then with probability 1 1 m rε,δ,m,Nε (· ∧ τε (δ, m, L)) , . . . , rε,δ,m,N (· ∧ τε (δ, m, L)) , . . . ∈ D∞,L . ε (14.29) Let Cˆ ∈ Rd and Dˆ ∈ Md×d arbitrary but fixed elements. Define for ε ≥ 0 maps Fˆε,i,L and σˆ ε,i,L from D∞,L × [0, ∞) into Rd and Md×d , respectively: For ε > 0, we set Fε (r i , Ψ Nε (r (·) ), t), if (·) ˆ Fε,i,δ,m,L (r , t) := ˆ C, if σε (r i , Ψ Nε (r (·) ), t), if σˆ ε,i,δ,m,L (r (·) , t) := ˆ D, if
⎫ ⎪ ⎪ ⎪ ⎪ (·) ⎪ r ∈ D∞,L , ⎪ ⎬ r (·) ∈ / D∞,L , ⎪ ⎪ ⎪ ⎪ r (·) ∈ D∞,L , ⎪ ⎪ ⎭ (·) r ∈ / D∞,L ,
(14.30)
and for ε = 0 we make corresponding definitions, employing in the above definition Ψ∞ . Stopping the solutions of (14.23) and (14.7) at τε (δ, m, L), we see by (14.27) and (14.29) that these stopped solutions solve the following systems of SODEs in (Rd )∞ : (·) (·) ⎫ i dˆrε,δ,m,N = Fˆε,i,δ,m,L rˆε,δ,m,Nε , t 1 D∞,L rˆε,δ,m,Nε dt ⎪ ε ⎪ ⎪ ⎬ (·) (·) i (14.31) +σˆ ε,i,δ,m,L (ˆrε,δ,m,Nε , t)1 D∞,L rˆε,δ,m,Nε dβε,δ,m, ⎪ ⎪ ⎪ ⎭ i rε,δ (0) = qεi , i = 1, . . . , Nε and
(·) (·) (·) (·) i dˆr0,∞ = Fˆ0,i,L rˆ0,∞ , t 1 D∞,L rˆ0,∞ dt + σˆ 0,i,L rˆ0,∞ , t 1 D∞,L rˆ0,∞ dβ i , i rε,δ (0) = qεi , i = 1, . . . , Nε , (14.32)
322
14 Partial Differential Equations as a Macroscopic Limit
respectively. Conversely, up to time τε (δ, m, L) every solution of (14.31) and (14.32) is a solution of (14.23) and (14.7), respectively. Note that i Fε r , Ψ N r (·) , t − F0 r i , Ψ∞ r (·) , t ≤ Fε r i , Ψ N r (·) , t ε ε −F0 r i , Ψ Nε r (·) , t + F0 r i , Ψ Nε r (·) , t − F0 r i , Ψ∞ r (·) , t . The last term in the above inequality tends to 0 by the Lipschitz assumption on F0 , as ε ↓ 0, if r (·) ∈ D∞ . The first terms tends to 0 by (14.8), as ε ↓ 0. Similarly, we can estimate the distance between the diffusion coefficients. Let K be a compact subset of Rd . Our assumption (14.8) now implies that for all i ≤ m, m ∈ N, δ > 0, and L ∈ N: lim sup sup sup Fˆε,i,δ,m,L r (·) , t − Fˆ0,i,δ,m,L r (·) , t ε↓0 r i ∈K 0≤t≤T r (·) (14.33) (·) (·) + σˆ ε,i,δ,m,L r , t − σˆ 0,i,δ,m,L r , t = 0 .
Consequently, Condition C.3 of Sect. 9 of Kurtz and Protter (loc. cit.) holds for (14.31) and (14.32). Finally, the Brownian motions in (14.23) and (14.6) are identically distributed (they are all standard). This implies that the sequence of integrators in (14.23) is uniformly tight (cf. Kurtz and Protter, loc. cit., Sect. 6). Hence, employing Theorem 9.4 of Kurtz and Protter (loc. cit.) in addition to (14.28) and the uniqueness of the solution of (14.7), we obtain that the solutions of (14.23) converge to the solutions of (14.7) for all m < Nε and all δ > 0.12 ⊔ ⊓
Let S be a complete, separable metric space. A family of Svalued random variables {ξ1 , . . . , ξm } is exchangeable if for every permutation (σ1 , . . . , σm ) of (1, . . . , m), {ξσ1 , . . . , ξσm } has the same distribution as {ξ1 , . . . , ξm }. A sequence ξ1 , ξ2 , . . . is exchangeable if every finite subfamily ξ1 , . . . , ξm is exchangeable. Let P(S) be the set of probability measures on (the Borel sets of) S. Lemma 14.6. For n = 1, 2, . . . , let {ξ1n , . . . , ξ Nn n } be exchangeable, Svalued random variables. (We allow Nn = ∞.) Let Ξ n be the corresponding empirical measure, Nn 1
n Ξ = δξ1n , Nn i=1
where if Nn = ∞, we mean
m 1
δξin . m→∞ m
Ξ n = lim
i=1
We will refer to the empirical process of an exchangeable system as the DeFinitti measure for the system. By the above convention this implies its continuum limit, if the system is infinite. 12
Cf. also the end of Sect. 15.2.5.
14.2 The Macroscopic Limit for d ≥ 2
323
Assume that Nn → ∞ and that for each m = 1, 2, . . . , {ξ1n , . . . , ξmn } ⇒ {ξ1 , . . . , ξm } in S m . Then {ξi } is exchangeable and setting ξin = s0 ∈ S for i > Nn , {Ξ n , ξ1n , ξ2n , . . .} ⇒ {Ξ, ξ1 , ξ2 , . . .} in P(S) × S ∞ ,
(14.34)
Ξ n → Ξ in probability in P(S).
(14.35)
where Ξ is the DeFinetti measure for {ξi }. If for each m, {ξ1n , . . . , ξmn } → {ξ1 , . . . , ξm } in probability in S m , then Proof. The exchangeability of {ξi } follows immediately from the exchangeability of {ξiNn }. Assuming m + k ≤ Nn , exchangeability implies / 0 n E f (ξ1n , . . . , ξm+k ) ⎡ 1 = E⎣ (Nn − m) · · · (Nn − m − k + 1) ⎤
n f ξ , . . . , ξmn , ξin , . . . , ξin ⎦ 1
=E
1
k
{i 1 ,...,i k }⊂{m+1,...,Nn }
Sk
n n 1 n n f ξ1 , . . . , ξm , s1 , . . . , sk Ξ (ds1 ) · · · Ξ (sk ) + O Nn
¯ m+k ), the bounded continuous realvalued functions on and hence if f ∈ C(S
S m+k ,
lim E f ξ1n , . . . , ξmn , s1 , . . . , sk Ξ n (ds1 ) · · · Ξ n (sk ) n→∞ k 0 / S = E f (ξ1 , . . . , ξm+k ) =E f (ξ1 , . . . , ξm , s1 , . . . , sk ) Ξ (ds1 ) · Ξ (dsk ) ,
Sk
where the second equality follows by exchangeability. Since the space of functions on P(S) × S ∞ of the form f (x1 , . . . , xm , s1 , . . . , sk )µ(ds1 ) · · · µ(dsk ) F(µ, x1 , . . . , xm ) = Sk
form a convergence determining class, the first part of the lemma follows. If for each m, {ξ1n , . . . , ξmn } → {ξ1 , . . . , ξm } in probability, then Ξn(m)
m m 1
1
n δξi → δξi ≡ m m i=1
i=1
in probability in P(S), and the convergence of Ξn to Ξ follows by approximation, ¯ that is, by exchangeability, for each ε > 0 and ϕ ∈ C(S), 5 B A B 4A lim sup P ϕ, Ξn(m) − ϕ, Ξn > ε = 0. m→∞ n
⊔ ⊓
324
14 Partial Differential Equations as a Macroscopic Limit
Corollary 14.7. Assuming the above hypotheses we obtain that for all m ∈ N and δ>0 (14.36) X ε,δ,m,Nε ⇒ X 0,∞ , as ε ↓ 0 . Proof. j
(i) Recall that, by assumption, {qεi }i∈N is exchangeable and P{qεi = qε } = 0 for i != j and ε ≥ 0. The assumptions on Fε and Jε as well as the usual properties of the conditional expectation and of {β i }i∈N imply Nε 1 (·)} is exchangeable. Denote by Ξε,δ,m,Nε (·), . . . , rε,δ,m,N that {rε,δ,m,N ε ε
Nε 1 (·)}, which is an ele(·), . . . , rε,δ,m,N the empirical process for {rε,δ,m,N ε ε ment of M1 (C([0, ∞); Rd )), i.e., of the space of probability measures on C([0, ∞); Rd ). The existence for Nε = ∞ follows from DeFinetti’s theorem (cf. Dawson (1993), Sect. 11.2). Let Ξ0,∞ be the DeFinetti measure for (14.7). By our Lemma 14.6,
Ξε,δ,m,Nε ⇒ Ξ0,∞ , as ε ↓ 0 .
(14.37)
Since X ε,δ,m,Nε (·) converges weakly, (14.36) follows.
⊔ ⊓
Recall the definition Λm := {( p 1 , . . . , p m ) ∈ Rd·m : ∃i != j, i, j ∈ {1, . . . , m}, with pi = p j }. 1 (·, q 1 ), . . . , Lemma 14.8. m is nonattainable by the Rd·m valued diffusion, (r0,∞ m m 1 m r0,∞ (·, q )) with initial values (q , . . . ., q ), which is defined by the first m ddimensional equations of the system (14.7).
Proof. Recall that we assumed d ≥ 2. It is obviously sufficient to prove the Lemma for m = 2, and, in what follows, we use a notation similar to Friedman (loc. cit.). We convert the nonautonomous equation into an autonomous one by adding time t as an additional dimension. Abbreviate F(r, t) := F0 (r, X 0,∞ (t), t), σ (r, t) := J0 (r, X 0,∞ (t), t) and set x := (r 1 , r 2 , t)T ∈ R2d+1 ,
where, as before, “AT ” denotes the transpose of a matrix A. Further, set ⎞ ⎛ ⎞ ⎛ 0 0 0 1 ˆ 0 ⎠, (14.38) F(x) := ⎝ F(r 1 , t) ⎠ , σˆ (x) = ⎝ 0 σ (r 1 , t) 2 0 0 σ (r 2 , t) F(r , t)
where σ is the appropriate d ×d matrix and σˆ (x) is a blockdiagonal matrix with entries in the first row and the first column all equal to 0. Let β(t) be a onedimensional standard Brownian motion, independent of the ddimensional standard Brownian motions β 1 (t) and β 2 (t) and set
14.2 The Macroscopic Limit for d ≥ 2
325
ˆ := (β(t), β 1 (t), β 2 (t))T . β(t) Then the stochastic SODEs for the two solutions of (14.7) can be written as an SODE in R2d+1 : ˆ ˆ dx = F(x)dt + σˆ (x)dβ, (14.39) x(0) = (0, q 1 , q 2 )T . Imbedded into R2d+1 , Λ2 becomes the ddimensional submanifold: Λˆ 2 = {(t, r 1 , r 2 ) : r 1 = r 2 }.
(14.40)
Normal vectors to Λˆ 2 are given by
1 N i := √ (0, . . . , 0, 1, 0, . . . , 0, −1, . . . , 0)T , 2 where the 1 is at the (i + 1)th coordinate and the −1 at the (i + d + 1)th coordinate and all other coordinates are 0, i = 1, . . . , d. The linear subspace, spanned by {N i , i = 1,. . . , d}, is ddimensional and will be denoted Λˆ ⊥ 2 . Further, let “•” denote here the scalar product in R2d+1 .Note that the diffusion matrix a(x) := σˆ (x)σˆ T (x) has a blockdiagonal structure, similar to σˆ (x). Indeed, for k ∈ {2, . . . , d + 1} and ℓ ∈ {d + 2, . . . , 2d + 1}, ⎫ aℓk (x) = akℓ (x) (by symmetry) ⎪ ⎪ ⎪ ⎪ ⎪ 2d+1 ⎪
⎪ ⎪ ⎪ = σˆ k j (x)σˆ ℓj (x) ⎪ ⎬ j=1 (∗) ⎪ 2d+1 d+1 ⎪
⎪ ⎪ = σˆ k j (x)σˆ ℓj (x) + σˆ k j (x)σˆ ℓj (x) ⎪ ⎪ ⎪ ⎪ ⎪ j=1 j=d+2 ⎪ ⎭ = 0 + 0, since σˆ k j (x) = 0 for k ≤ d + 1 and j ≥ d + 2 and σˆ ℓj (x) = 0 for ℓ ≥ d + 2 and j ≤ d + 1. Further, since σˆ 1 j (x) = σˆ i1 (x) = 0 ∀i, j, we have akℓ (x) = 0 if k = 0 or ℓ = 0.
So (a(x)N i ) · N j =
d √
σ jk (r 1 , t)σik (r 1 , t), 2 k=1
(∗∗)
if x = (t, r 1 , r 1 )T ∈ Λˆ 2 . (14.41)
Since σ ji (r, t) is invertible for all (r, t), we obtain that for x ∈ Λˆ 2 the rank of a(x), restricted to Λˆ ⊥ 2 , equals d. The statement of the Lemma now follows from Friedman (loc. cit., Chap. 11, (1.8) and Theorem 4.2). ⊔ ⊓
326
14 Partial Differential Equations as a Macroscopic Limit
Lemma 14.9. Assuming the above hypotheses we obtain that for all t ≥ 0 there is a sequence δm := δm (t) ↓ 0 as m → ∞ such that lim limε↓0 P{τεm (δm ) ≤ t} = 0.
m→∞
(14.42)
Proof. For δ ≥ 0, set 4 5 Λm (δ) := ( p 1 , . . . , p m ) ∈ Rd·m : ∃i != j, i, j ∈ {1, . . . , m}, with  pi − p j  ≤ δ (14.43) and note that Λm (0) = Λm , where Λm was defined before Theorem 14.2. Define maps τ m (δ) from CRdm [0, ∞) into [0, ∞) through τ m (δ)((r 1 (·), . . . , r m (·))) := inf{t : (r 1 (t), . . . , r m (t)) ∈ m (δ)}. Observe that 4 1 5 (r (·), . . . , r m (·)) : τ m (δ)((r 1 (·), . . . , r m (·))) ≤ t 5 4 = (r 1 (·), . . . , r m (·)) : ∃s ∈ [0, t], i != j such that r i (s) − r j (s) ≤ δ ,
and we verify that the latter set is closed in C([0, ∞); Rdm ), which is endowed with the metric of uniform convergence on compact intervals [0, T ]. Setting 5 4 i j (t) − r0,∞ (t) ≤ δ , τ0m (δ) := min inf t : r0,∞ 1≤i!= j≤m
we have
1 m τ m (δ)((r0,∞ (·, ω), . . . , r0,∞ (·, ω)))
= τ0m (δ, ω) and for δ > 0 τ m (δ)((rε1 (·, , ω), . . . , rεm (·, ω))) = τεm (δ, ω) . Hence, by a standard theorem in weak convergence,13 for all δ > 0 and m < Nε 4 5 4 5 limε↓0 P τεm (δ) ≤ t ≤ P τ0m (δ) ≤ t . (14.44) By Lemma 14.8 for all m
4 5 lim P τ0m (δ) ≤ t = 0 ∀t > 0. δ↓0
Hence there is a sequence δm = δm (t) ↓ 0, as m → ∞ such that 4 5 lim P τ0m (δm ) ≤ t = 0. m→∞
Expressions (14.44) and (14.46) together imply (14.42). 13
Cf. Theorem 15.24, Sect. 15.2.1.
(14.45)
(14.46) ⊔ ⊓
14.3 Examples
327
Note that w.p. 1 0 / i rε,δ,N (t) ≡ rεi (t) on 0, τεm (δ) , i = 1, . . . , m ≤ Nε , ε
(14.47)
where {rεi (t)}i=1,...,Nε is the solution of (4.10) with σi⊥ ≡ 0 ∀i. Proof of Theorem 14.2
Employing (14.42), Lemma 14.5, and Corollary 14.7, we obtain X ε,Nε , rε (·, X ε,Nε , qε1 ), . . . , rε (·, X ε,Nε , qεNε ) ⇒ (X 0,∞ , r0,∞ (·, X 0,∞ , q 1 ), . . .) in C([0, ∞); M × (Rd )∞ ), as ε → 0 .
(14.48)
Since the solution X 0,∞ of (14.6) is unique (cf. Remark 14.1), (14.14) follows from Lemma 14.3. The proof of (14.15) is easier. ⊔ ⊓
14.3 Examples We provide examples of kernels, satisfying the additional hypotheses from Sect. 5.1. (i) Fε (r, µ) ≡ F0 (r, µ) for ε ≥ 0. Let α > 0. Set d/4 2 exp − r4α and Γα (r ) the diagonal d × d matrix, whose Γ˜α (r ) := (2π1 α) entries on the main diagonal are all Γ˜ε (r ). Set Jε (r, p, µ) := Γε (r − p)Γ2 ( p − q)µ(dq). Then, D˜ ε (µ, r, q) =
Γε (r − p)Γε (q − p)Γ2 ( p − q)Γ ˜ 2 ( p − q)µ(d ˆ q)µ(d ˜ q)d ˆ p.
Since we are dealing with diagonal matrices with identical entries we may, in what follows, assume d = 1. Then, Γε2 (r − p)Γ2 ( p − q)Γ ˜ 2 ( p − q)µ(d ˆ q)µ(d ˜ q)dp ˆ σε2 (r, µ) = Dε (µ, r ) =
Γε2 (r − p)d p = 1, σ02 (r, µ) = Γ2 (r − q)µ(d ˜ q)Γ ˜ 2 (r − q)µ(d ˆ q) ˆ = Γε2 (r − p)Γ2 (r − q)µ(d ˜ q)Γ ˜ 2 (r − q)µ(d ˆ q)d ˆ p.
and, since
328
14 Partial Differential Equations as a Macroscopic Limit
We employ the Chapman–Kolmogorov equation to “expand” Γ2 (q − q) ¯ = c1 Γ1 (q − p)Γ ¯ 1 ( p¯ − q)d ¯ p) ¯ with sup Γ1 (q) ≤ c2 < ∞. Hence, q
Γ2 ( p − q)µ(d ˜ q)Γ ˜ 2 ( p − q)µ(d ˆ q) ˆ − Γ2 (r − q)µ(d ˜ q)Γ ˜ 2 (r − q)µ(d ˆ q) ˆ / 0 2 Γ1 ( p − p)Γ = c1 ˜ 1 ( p − p) ˆ − Γ1 (r − p)Γ ˜ 1 (r − p) ˆ ˜ q)Γ ˜ 1 ( pˆ − q)µ(d ˆ q) ˆ d p˜ d pˆ Γ1 ( p˜ − q)µ(d ≤ c12 c22 Γ1 ( p − p)Γ ˜ 1 ( p − p) ˆ − Γ1 (r − p)Γ ˜ 1 (r − p)d ˆ p˜ d p, ˆ where we used µ(Rd ) = 1. So, σε2 (r, µ) − σ02 (r, µ) ≤ c Γε2 (r − p)Γ1 ( p − p)Γ ˜ 1 ( p − p) ˆ − Γ1 (r − p)Γ ˜ 1 (r − p)d ˆ p˜ d pˆ d p = c Γε2 ( p) f (r, p)d p
by change of variables with f (r, p) := Γ1 (r + p − p)Γ ˜ 1 (r + p − p) ˆ − Γ1 (r − p)Γ ˜ 1 (r − p)d ˆ p˜ d p. ˆ f is continuous in R2d with f (r, 0) = 0 ∀r . Since Γε2 ( p) −→ δ0 as ε ↓ 0, we have for a compact set K ⊂ Rd sup Γε2 ( p) f (r, p)d p → 0, as ε ↓ 0. r ∈K
This implies (14.8). (ii) Let µ ∈ C, where C is a compact subset of M1 . By Prohorov’s theorem for any δ > 0 there is an L > 0 such that infµ∈C µ(SL ) ≥ 1 − δ. Let SL := { p ∈ Rd ;  p ≤ L}. Set Γ2 ( p − q). cδ := inf p∈SL ,q∈S ˜ L
Obviously, c L > 0. So, σε2 (r, µ) = σε2 (r, µ)
=
≥ ≥
Γε2 (r − p) Γε2 (r
SL
SL
− p)
Γε2 (r Γε2 (r
− p) −
2 ˜ q) ˜ dp Γ2 ( p − q)µ(d
2 ˜ q) ˜ dp Γ2 ( p − q)µ(d
SL
2 Γ2 ( p − q)µ(d ˜ q) ˜ dp
p)cδ2 [1 − δ]2 d p.
14.3 Examples
329
Since {r } is bounded in the formulation of assumption in (14.9), we may assume r  ≤ L2 . Changing variables in SL Γε2 (r − p)d p we obtain that for r  ≤ L2 and 0 < ε ≤ 1, Γε2 (r − p)d p ≥ Γ12 ( p)d p ≥ c L > 0. { p≤ L2 }
SL
Altogether we obtain (14.9). (iii) Set c2 := sup Γ2 (q), q
which is obviously finite. Since µ(Rd ) = 1, we obtain Dε (r, q, µ) = ≤
c22
Γε (r − p)Γε (q − p)
2 Γ2 ( p − q)µ(d ˜ q) ˜ dp
Γε (r − p)Γε (q − p)d p =
c22 exp
r − q2 , − 8ε
whence we obtain (14.10). (iv) We are checking the Lipschitz condition (4.11) for our example, assuming again without loss of generality d = 1. Jε (r, p, µ) := Γε (r − p)Γ2 ( p − q)µ(dq). (Jε (r1 , p, µ1 ) − Jε (r2 , p, µ1 ))2 d p 32 [Γε (r1 − p) − Γε (r2 − p)] Γ2 ( p − q)µ(dq) d p = r1 − r2 2 ≤ c22 [Γε (r1 − p) − Γε (r2 − p)]2 d p = 2c22 1 − exp − 8ε 2 r − r  1 2 ∧ 1 ≤ cε ρ 2 (r1 − r2 ). ≤ 2c22 8ε Γ2 ( p −·) is a bounded function of q with bounded derivatives, where the bound can be taken uniform in p. Therefore, Γ2 ( p − q1 ) − Γ2 ( p − q2 ) ≤ c2 ρ(q1 − q2 ). Hence,
(Jε (r, p, µ1 ) − Jε (r, p, µ2 ))2 d p 32 Γε (r − p) Γ2 ( p − q)[µ1 (dq) − µ2 (dq)] d p = 2 2 ≤ c γ (µ1 − µ2 ) Γε2 (r − p)d p = c2 γ 2 (µ1 − µ2 ).
330
14 Partial Differential Equations as a Macroscopic Limit
(v) Assume, without loss of generality, F ≡ 0. We have D0 (r, t) := Γ2 (r − q)X ˜ 0,∞ (t, dq)Γ ˜ 2 (r − q)X ˆ 0,∞ (t, dq). ˆ Obviously, D0 (r, t) is infinitely often differentiable with respect to r with derivative continuous in (r, t). Note that the “test functions” Γ2 (r − q) are in the Schwarz space of infinitely often differentiable functions, with all derivatives rapidly decreasing. Taking all derivatives in the integrals in the distributional sense, ∂ ∂ Γ2 (r − q) ˜ ˜ 2 (r − q)X ˆ 0,∞ (t, dq). ˆ D0 (r, t) = 2 X 0,∞ (t, dq)Γ ∂t ∂t Further, by (14.6), we may replace the partial with respect to t by the (quasilinear) second partial with respect to the space variable, i.e., ∂ D0 (r, t) = Γ2 (r − q)∂ ˜ q˜ q˜ [D0 (q, ˜ t)X 0,∞ (t, dq)]Γ ˜ 2 (r − q)X ˆ 0,∞ (t, dq). ˆ ∂t By definition (“integrating by parts”) the righthand side can be rewritten, taking the second derivative of the “test function” Γ2 (r − q), ˜ with respect to q. ˜ By the homogeneity of Γ2 (r − q), ˜ this second derivative equals the second derivative with respect to r . Hence, ∂ D0 (r, t) = (∂rr Γ2 )(r − q)D ˜ 0 (q, ˜ t)X 0,∞ (t, dq)Γ ˜ 2 (r − q)X ˆ 0,∞ (t, dq). ˆ ∂t We verify that ∂t∂ D0 (r, t) is continuously differentiable in all variables. Hence, we may compute the second derivative with respect to t, obtaining two terms. One of them contains ∂t∂ D0 (r, t) as a factor in the integral which we may replace by the righthand side of the last equation. The other term contains ∂ ˜ as a factor in the integral. As before, we replace this term by ∂t X 0,∞ (t, dq) ∂q˜ q˜ [D0 (q, ˜ t)X 0,∞ (t, dq)] ˜ and integrate by parts. Altogether, we obtain that D0 (r, t) is twice continuously differentiable with respect to all variables. The statement for d ×d matrixvalued coefficients follows from the onedimensional case.
14.4 A Remark on d = 1 Based on a result by Dorogovtsev (2005) we expect in the onedimensional case the solution of (8.26) not to converge to the solution of a macroscopic PDE but to the solution of an SPDE driven by one standard Brownian motion β(t) instead of the space–time white noise w(dq, dt). A rigorous derivation of this SPDE is planned for a future research project.
14.5 Convergence of Stochastic Transport Equations to Macroscopic Parabolic Equations
331
14.5 Convergence of Stochastic Transport Equations to Macroscopic Parabolic Equations We obtain that a sequence of firstorder stochastic transport equations, driven by Stratonovich differentials converges to the solution of a semilinear parabolic PDE, as the correlation length tend to 0. We reformulate Theorem 14.2, employing the representation of the solutions of (8.26) via a firstorder stochastic transport equation, driven by Stratonovich differentials. Theorem 14.3. Suppose the conditions of Theorem 14.2 with diffusion kernel Jε (r, q, u) being independent of the measure variables. Consider the sequence of solutions of dX ε = − ▽ ·(X ε F(·, X ε , t))dt − ▽ · X ε Jε (·, p, t) w(d p, ◦dt), (14.49) X ε (0) = X ε,0 . Then X ε ⇒ X 0,∞ in C([0, ∞); M1 ), as ε ↓ 0 ,
(14.50)
where X 0,∞ (·) is the solution of the semilinear McKean–Vlasov equation ⎫ d ⎪ ∂ 1 2 ⎬ X 0,∞ = ∂kℓ (D0,kℓ (·, t)X 0,∞ ) − ▽ · (X 0,∞ F0 (·, X 0,∞ , t)), (14.51) ∂t 2 k,ℓ=1 ⎪ ⎭ X 0,∞ (0) = µ.
Proof. Theorem 14.2 in addition to (8.81).
⊔ ⊓
Chapter 15
Appendix
15.1 Analysis 15.1.1 Metric Spaces: Extension by Continuity, Contraction Mappings, and Uniform Boundedness The proof of the following statement may be found in Dunford and Schwartz ((1958), Sect. I.6., Theorem 17). Proposition 15.1. Principle of Extension by Continuity Let B1 and B2 be metric spaces and let B2 be complete. If f : A %→ B2 is uniformly continuous on the dense subset A of B1 , then f has a unique continuous extension f¯ : B1 %→ B2 . This unique extension is uniformly continuous on B1 .
⊔ ⊓
The following theorem is also known as “Banach’s fixed point theorem.” For the proof we refer to Kantorovich and Akilov (1977), Chap. XVI.1. Theorem 15.2. Contraction Mapping Principle Let (B, dB ) be a complete metric space and B a closed subset of B. Suppose there is a mapping :B→B such that there is a δ ∈ (0, 1) and the following holds: dB (( f ), (g)) ≤ δdB ( f, g) ∀ f, g ∈ B. 335
336
15 Appendix
Then there is a unique f ∗ ∈ B such that1
( f ∗ ) = f ∗ .
⊔ ⊓
The following uniform boundedness principle and the preceding definitions are found in Dunford and Schwartz (1958), Sect. II, Theorem 11, and Definitions 7 and 10 and Yosida (1968), Theorem 1 and Corollary 1 (Resonance Theorem). First, we require some terminology Let (B, dB ) be a metric space. The metric dB ( f, g) is called a “Fr´echet metric” (or “quasinorm”) if dB ( f, g) = dB ( f − g, 0) ∀ f, g ∈ M. • If the metric space B, endowed with a Fr´echet metric, dB is also linear topological space2 it is called a “quasinormed linear space.” • If a quasinormed space B is complete it is called a “Fr´echet Space.” • A subset B of a quasinormed space B is “bounded” if for any ε > 0 there is a γ > 0 such that dB (γ B, 0) ≤ ε. Theorem 15.3. Uniform Boundedness Principle (I) For each a in a set A, let Ta be a continuous linear operator from a Fr´echet space B1 into a quasinormed space B2 . If, for each f ∈ B1 , the set {Ta f : a ∈ A} is bounded in B2 , then lim Ta f = 0 uniformly for a ∈ A .
f →0
(II) If, in addition to the assumptions of Part (I), B1 is a Banach space with norm ·B1 and B2 is a normed vector space with norm ·B2 , then the boundedness of {Ta f B2 : a ∈ A} at each f ∈ B1 implies the boundedness of {Ta L(B1 ,B2 ) } where Ta L(B1 ,B2 ) := sup Ta f B2 f B1 ≤1
⊔ ⊓
is the usual operator norm.
15.1.2 Some Classical Inequalities ˜ F, µ) be some measure space and for p ∈ [1, ∞) let L p (, ˜ F, µ) the space Let (, of measurable real (or complex)valued functions f such that 31 p < ∞.  f  p (ω)µ(dω) ˜ f ˜ p := ˜ 
1 2
The mapping is called a “contraction” and f ∗ a “fixed point.” This means that (B, dB ) is a vector space such that addition of vectors and multiplication by scalars are continuous operations.
15.1 Analysis
337
Proposition 15.4. H¨older’s Inequality 3 p ˜ Then Let p ∈ (1, ∞) and p¯ := p−1 and f and g be measurable functions on . ˜ ˜ p¯ . ˜ f g ˜ 1 ≤ ˜ f ˜ p g
(15.1)
Proof. 4 (i) We first show ab ≤
ap b p¯ + p p¯
for nonnegative numbers a and b and p > 1 and p¯ :=
p p−1 .
(15.2) tp p
Indeed, the function h(t) :=
1 p¯
+ − t has for t ≥ 0 the minimum value 0, and − p¯
this minimum is attained only at t = 1. Setting t = ab p , we obtain (15.2). ˜ ˜ ¯ = 0 (since then f = 0 µ−a.e. or (ii) The result is trivial if ˜ f ˜ p = 0 or g p ˜ ˜ p¯ = ∞. Otherwise, we g = 0 µ−a.e.) It is also trivial if ˜ f ˜ p = ∞ or g apply (15.2) with g(ω)  f (ω) a := , b := . ˜ f ˜ p ˜ ˜ p¯ g Then
 f  p (ω) g p¯ (ω)  f (ω)g(ω) ≤ . p + ˜ ˜ p¯ ˜ ˜ p¯ p ˜ f ˜ p ˜ f ˜ p g p¯ g p¯
Integrating both sides yields ˜ f g ˜ 1 1 1 ≤ + . ˜ ˜ p¯ p p¯ ˜ f ˜ p g ⊔ ⊓ Next, we derive Gronwall’s inequality.5 Proposition 15.5. Gronwall’s Inequality Let ϕ, ψ, χ be realvalued continuous (or piecewise continuous) functions on a real interval a ≤ t ≤ b. Let χ (t) > 0 on [a, b], and suppose that for t ∈ [a.b] t ϕ(t) ≤ ψ(t) + χ (s)ϕ(s)ds. (15.3) a
Then ϕ(t) ≤ ψ(t) + 3 4 5
a
t
χ (s)ψ(s)ds exp
s
t
χ (u)du ds.
(15.4)
For p = p¯ = 2 we have the CauchySchwarz inequality. We follow the proofs, provided by Adams (1975), Chap. II, Theorem 2.3, and Folland (1984), Sect. 6, Theorem 6.2. Cf. Coddington and Levinson (1955), Chap. I.8, Problem 1.
338
15 Appendix
Proof. Set x(t) :=
t a
χ (s)ϕ(s)ds. We immediately verify that (15.3) implies
x ′ (t) − χ (t)x(t) ≤ χ (t)ψ(t) ∀t ∈ [a, b], x(a) = 0. Set
f (t) := x ′ (t) − χ (t)x(t) ∀t ∈ [a, b].
Obviously, x(·) solves the ODE initial value problem x ′ (t) = χ (t)x(t) + f (t), x(a) = 0. Employing variation of constants, x(·) has the following representation t t x(t) = exp χ (u)du f (s)ds. a
s
Since f (t) ≤ χ (t)ψ(t) ∀t ∈ [a, b] we obtain t t x(t) ≤ exp χ (u)du χ (s)ψ(s)ds ∀t ∈ [a, b], a
s
whence ′
χ (t)ϕ(t) = x (t) ≤ χ (t)
a
t
exp
s
t
χ (u)du χ (s)ψ(s)ds+χ (t)ψ(t) ∀t ∈ [a, b].
Dividing both sides of the last inequality by χ (t) yields (15.4).
⊔ ⊓
Let (1 , F1 , µ1 ) and (2 , F2 , µ2 ) be σ finite measure spaces. L p ((Bi , Fi , µi )) are the spaces of pintegrable functions on the measure space i , and the corresponding L p norms will be denoted f i , p , i = 1, 2, where 1 ≤ p ≤ ∞. Proposition 15.6. Inequality for Integral Operators Let K be an F1 ⊗ F2 measurable function on 1 × 2 and suppose there exists a finite C > 0 such that K (ω1 , ω2 )µ(dω1 ) ≤ C µ2 a.e. and K (ω1 , ω2 )µ(dω2 ) ≤ C µ1 a.e.
(15.5)
Then, f ∈ L p ((2 , F2 , µ2 )) for 1 ≤ p ≤ ∞ implies that ¯ ( K f )(ω1 ) := K (ω1 , ω2 ) f (ω2 )µ2 (dω2 ) ∈ L p ((1 , F1 , µ1 )),
(15.6)
such that K¯ f 1 , p ≤ C f 2 , p .
(15.7)
15.1 Analysis
339
Proof. 6 Suppose p ∈ (1, ∞). Then K (ω1 , ω2 ) f (ω2 )µ2 (dω2 ) p−1 1 + = K (ω1 , ω2 ) p p  f (ω2 )µ2 (dω2 ) 1 p−1 p p p ≤ K (ω1 , ω2 ) f (ω2 ) µ2 (dω2 ) K (ω1 , ω2 )µ2 (dω2 ) ≤C
(by H¨older’s inequality (Proposition 15.4)) 1 p p µ1 a.e. K (ω1 , ω2 ) f (ω2 ) µ2 (dω2 )
p−1 p
Hence, by the Fubini–Tonelli theorem, p K (ω1 , ω2 ) f (ω2 )µ2 (dω2 ) µ1 (dω1 ) p−1 ≤C K (ω1 , ω2 ) f (ω2 ) p µ2 (dω2 )µ1 (dω1 ) ≤ C p  f (ω2 ) p µ2 (dω2 ) < ∞. We obtain from the last estimate that, by Fubini’s theorem, K (ω1 , ·) f (·) ∈ L 1 ((2 , F2 , µ2 )) µ1 a.e.. Therefore, K¯ f is welldefined µ1 a.e., and p ( K¯ f )(ω1 ) p µ1 (dω1 ) ≤ C p f 2 , p . Taking the pth root on both sides, we obtain (15.7). For p = 1 we do not require H¨older’s inequality, and other than that the proof is similar. For p = ∞ the proof is trivial. ⊔ ⊓ Proposition 15.6 immediately implies the classical Young inequality.7 Proposition 15.7. Young’s Inequality Let f ∈ W0,1,1 and g ∈ W0, p,1 , where 1 ≤ p ≤ ∞. Then the convolution f ∗ g exists for almost every r ∈ Rd such that f ∗ g ∈ W0, p,1 and f ∗ g0, p,1 ≤ f 0,1,1 g0, p,1 . Proof. We set K (r, q) := f (r − q) and apply Proposition 15.6.8 6 7 8
(15.8) ⊔ ⊓
The proof has been adopted from Folland (1984), Sect. 6.3, (6.18) Theorem. We remark that in Tanabe (1979), Sect. 1.2, Lemma 1.2.3, this inequality is called the “HaussdorfYoung inequality.” Cf. Folland (loc.cit), Sect. 8.1, (8.7).
340
15 Appendix
15.1.3 The Schwarz Space We define the system of normalized Hermite functions on H0 := L 2 (Rd , dr ) and provide a proof of the completeness of the system. Further, we prove that the normalized Hermite functions are eigenfunctions of −△ +  · 2 , where  · 2 acts as a multiplication operator. We then introduce the Schwarz space and its strong dual as well as the chain of associated Hilbert spaces. Finally, we define those Hilbert distribution spaces, H−γ , for which the imbedding H0 ⊂ H−γ is HilbertSchmidt. As in Chap. 3 we use the abbreviation ˆ := N ∪ {0}. N First, we define the usual Hermite functions (i.e,. the Hermite functions that are not normalized!) and the normalized Hermite functions for d = 1 and then for d > 1: (i) Case d = 1.
⎫ x 2 dk 2 ⎪ gk (x) := (−1) exp exp(−x ), ⎪ ⎬ 2 dx k 1 ⎪ ˆ ⎪ φk := 1 ⎭ √ gk , k ∈ N. k 2 k! π k
(ii) Case d > 1.
(15.9)
Denote the multiindices with a single letter in bold face, i.e., set k := (k1 , . . . , kd ) k := k1 + . . . + kd . Define: gk (q1 , . . . , qd ) :=
d
gki (qi ),
i=1
ˆ i = 1, . . . , d. ki ∈ N,
(15.10)
We then obtain the normalized Hermite functions9 d
1 Φk := 6 φk i . gk = 7 d 7 k i=1 8 2 i ki !π d4
(15.11)
i=1
Further, let △ denote the selfadjoint closure of the Laplace operator on H0 with Dom(△) = H2 .  · 2 denotes the multiplication operator acting on functions f from a suitable subspace of H0 . 9
Cf. (15.28) in the proof of the following Proposition 15.8 which establishes that the normalized Hermite functions have norm 1 in H0 .
15.1 Analysis
341
Proposition 15.8. The system {φk } is a complete orthonormal system for H0 = L 2 (Rd , dr ). Further, ((−△ + r 2 )φk )(r ) ≡ (2k + d)φk (r ) ∀k, (15.12)
i.e., the φk are eigenfunctions of the selfadjoint unbounded operator −△ +  · 2 on H0 .10 Proof. (i) Suppose first d = 1. To show the orthogonality we set ∞ Jk,ℓ := gk (x)gℓ (x)dx −∞
and suppose k ≥ ℓ. Define the ℓth Hermite polynomial by h ℓ (x) := (−1)ℓ exp(x 2 ) Observe that h ℓ (x) exp and we verify that
x2 − 2
dℓ exp(−x 2 ) dx ℓ
(15.13)
≡ gℓ (x),
h ℓ (x) = 2ℓ x ℓ + pℓ−2 (x),
(15.14)
where pℓ−2 (x) is a polynomial of degree ℓ − 2. Therefore, k ℓ ∞ d d k+ℓ 2 2 2 exp(x ) exp(−x ) exp(−x ) dx Jk,ℓ = (−1) dx k dx ℓ −∞ ∞ k d exp(−x 2 )h ℓ (x)dx. = (−1)k k −∞ dx Integrating by parts k times we obtain k ∞ d 2k 2 Jk,ℓ = (−1) exp(−x ) h (x)dx. k ℓ dx −∞ Observing that the degree of h ℓ (·) is ℓ we have k d 2k k!, if ℓ = k, (x) := h ℓ k dx 0, if ℓ < k. Hence,
√ 2k k! π , if ℓ = k, (15.15) 0, if ℓ < k. Equation (15.15) establishes the orthogonality of the system {gk (·)} and the orthonormality of the system {φk (·)}. Jk,ℓ :=
10
Most of the steps in the following proof are taken from Suetin (1979), Chap. V, Sects. 1 and 3 as well as from Akhiezer and Glasman (1950), Chap. 1, Sect. 11. Cf. also the appendix in the paper by Holley and Stroock (1978) and Kotelenez (1985).
342
15 Appendix
(ii) Next, we show completeness of the onedimensional system. Suppose there is an f ∈ L 2 (R, dx) such that ∞ ˆ f (x)φk (x)dx = 0 ∀k ∈ N.
(15.16)
−∞ 2
ˆ we observe that, for any n ∈ N, ˆ the linSetting φ˜ k (x) := exp(− x2 )x k , k ∈ N, ˜ ˜ ear spans of {φ0 , . . . , φn } and of {φ0 , . . . , φn } coincide. Thus, the assumption (15.16) is equivalent to 2 ∞ x ˆ (15.17) x k dx = 0 ∀k ∈ N. f (x) exp − 2 −∞ √ Setting i := −1 and for z ∈ C 2 ∞ x (15.18) ψ(z) := f (x) exp − exp(ix z)dx, 2 −∞ we obtain (employing Lebesgue’s dominated convergence theorem) that ψ(·) is analytic in the complex plane. By (15.17) Taylor’s expansion at z = 0 yields 2 ∞ x ˆ (15.19) f (x) exp − (ix)k dx = 0 ∀k ∈ N, ψ (k) (0) = 2 −∞ where ψ (k) (0) are the kth derivatives of ψ at z = 0. Hence, ψ(z) ≡ 0. In particular, we have ψ(ξ ) =
∞
−∞
2 x exp(ixξ )dx = 0 ∀ξ ∈ R. f (x) exp − 2
(15.20)
Multiplying both sides by exp(−iξ y), y ∈ R, and integrating with respect to dξ from −γ to γ , yields 2 ∞ sin(γ (x − y)) x (15.21) dx = 0 ∀γ and ∀y . f (x) exp − 2 x−y −∞ (x−y)) can be extended to a bounded and continuous function For fixed γ sin(γx−y in both variables, where the value at 0 equals γ . Let −∞ < a < b < ∞ and integrate (15.21) with respect to dy from a to b. Employing the Fubini–Tonelli theorem we obtain from (15.21) 2 b ∞ x sin(γ (x − y)) (15.22) f (x) exp − dy dx = 0 ∀γ . 2 x−y −∞ a
15.1 Analysis
343
The following relation is well known and proven in classical analysis books:11 ∞ sin(u) π du = . (15.23) u 2 0 Let c > 0. Then
0
c
∞
sin(u) du = u
c∧((k+1)π )
k=0 c∧(kπ )
sin(u) du. u
The series on the right is alternating, whose terms decrease to 0 in absolute value. The first term is in (0, π). Therefore, c sin(u) du ∈ (0, 2π ) ∀c > 0. u 0 Since
sin(u) is even, u b sin(γ (x − y)) dy ≤ 4π ∀γ and ∀x x−y a
Changing variables, b γ (b−x) sin(γ (x − y)) sin(γ (u)) dy = du. x − y u a γ (a−x) Therefore, by (15.23), b π sin(γ (x − y)) lim dy = 1(a,b) (x)π + 1{a,b} (x) , γ →∞ a x−y 2
(15.24)
where 1{a,b} (x) is the indicator function of the set {a, b}, containing the two elements a, b. By Lebesgue’s dominated convergence theorem, 2 2 b b ∞ x x sin(γ (x − y)) dy dx = π f (x) exp − dx. lim f (x) exp − γ →∞ −∞ 2 x − y 2 a a Hence, by (15.22), 2 b x dx = 0 , −∞ < a < b < ∞. f (x) exp − 2 a
(15.25)
The integral in (15.25) is an absolutely continuous function of b, whence 2 f (x) exp − x2 = 0 dx a.e., which is equivalent to f (x) = 0 dx a.e. (15.26) This establishes the completeness of the system {φk } for d = 1. 11
Cf., e.g., Erwe (1968), Chap. VI.8, (289).
344
15 Appendix
(iii) The completeness of the system {φk } in H0 = L 2 (Rd , dr ) follows from the completeness of the onedimensional system employing in the following relation Fubini’s and the monotone convergence theorems: ∞ ∞ ∞ ⎫ ⎪ ⎪ f 2 (r1 , . . . , rd )dr1 dr2 · · · drd ···  f 20 = ⎪ ⎪ ⎪ −∞ −∞ −∞ ⎪ ⎡ ⎤ ⎪ ⎪ 2 ∞ ∞ ∞ ⎪ ⎪ ⎪ ⎪ ⎣ ⎪ f (r1 , . . . , rd )φk1 (r1 )dr1 ⎦dr2 · · · drd . . . .. = ⎪ ⎪ −∞ −∞ −∞ ⎪ k1 ⎪ ⎪ ⎪ ⎪ ⎪ ∞ ∞
∞ ⎪
∞ ⎪ ⎬ f (r1 , . . . , rd )φk (r1 )dr1 = ··· k1
−∞
−∞
.. . =
k1
....
kd
∞ −∞
−∞
k2
−∞
1
⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ 2 ⎪ ⎪ ⎪ ⎪ f (r1 , . . . , rd )φk1 (r1 ) · · · · · φkd (rd )dr1 · · · drd .⎪ ⎪ ⎭
2 dr3 · · · drd φk2 (r2 )dr2
···
∞
−∞
(15.27)
Further, the orthonormality follows from the onedimensional case where (15.15) provides the correct normalizing factor also for the ddimensional case, i.e., 6 7 d 7 d gk 0 = 8 2ki ki !π 4 ∀k. (15.28) i=1
(iv) We next prove (15.12) for d = 1. Observe that
d exp(−x 2 ) ≡ −2x exp(−x 2 ), dx whence we recursively establish dk+2 dk+1 dk 2 2 exp(−x 2 ) ≡ 0. exp(−x ) + 2x exp(−x ) + 2(k + 1) dx k dx k+2 dx k+1 Recalling (15.13), the last equation may be rewritten as d2 d 2 2 2 exp(−x exp(−x )h (x) +2(k +1) exp(−x )h (x) ≡ 0. )h (x) +2x k k k dx dx 2 Abbreviating f (x) := exp(−x 2 ) we obtain from the previous equation and the equations for the first and second derivatives of f
15.1 Analysis
345
d2 d f (x) + 2(k + 1) f (x) h k (x) f (x) + 2x dx dx 2 d d d2 + 2 f (x) + 2x f (x) h k (x) + f (x) 2 h k (x) dx dx dx 2 d d = f (x) h k (x) − 2x h k (x) + 2kh k (x) ≡ 0 dx dx 2
Dividing the last equation by f (x) = exp(−x 2 ) yields the differential equations for the Hermite polynomials 2 d d (x) + 2kh k (x) ≡ 0. (x) − 2x (15.29) h h k k dx dx 2 x2 We differentiate gk (x) = exp − h k (x) and obtain 2 2 2 d d x x h k (x) + exp − gk (x) = −x exp − h k (x) dx 2 2 dx and 2 d2 d2 d x 2 h −h (x) + (x) + x h (x) − 2x g (x) = exp − h (x) . k k k k k 2 dx dx 2 dx 2 Thus, d2 gk (x) + (2k + 1 − x 2 )gk (x) dx 2 2 2 d x d ≡ exp − h k (x) − 2x h k (x) 2 2 dx dx +x 2 h
k (x) − h k (x) + (2k
+1−
x 2 )h
k (x)
%
2 2 x d d ≡ exp − h (x) + 2kh (x) ≡0 h (x) − 2x k k k 2 dx dx 2 by (15.29). As a result we obtain the following differential equation for the Hermite functions: 2 d (15.30) gk (x) + (2k + 1 − x 2 )gk (x) ≡ 0. dx 2 Obviously, the normalized Hermite functions Φk (·) are solutions of the same differential equation (15.30), whence we obtain (15.12) for d = 1. d ∂ii2 and (15.30) Expression (15.11) in addition to the definition of △ = i=1
proves (15.12) also for d > 1.
⊔ ⊓
346
15 Appendix
Let us introduce suitable spaces of distributions. The following setup is standard and details can be found in Itˆo (1984), Kotelenez (1985), and Holley and Stroock (loc.cit) and the references therein. For γ ≥ 0 and f, g ∈ Cc∞ (Rd ; R) (the infinitely often continuously differentiable real valued functions with compact support in Rd ) we define scalar products and Hilbert norms by
⎫ & f, g'γ := (d + k)γ & f, φk '0 &g, φk '0 ; ⎬ (15.31) ˆd k∈N 1 ⎭ f γ := & f, f 'γ . Define Hilbert spaces for γ ≥ 0 by
Hγ := { f ∈ H0 : f γ < ∞}, where H0 = H0 . Set
S := ∩γ ≥0 Hγ .
ˆ This The topology on S is defined by the countable Hilbert norms · γ , γ ∈ N. topology is equivalent to the topology generated by the metric dS,H ( f, g) :=
∞
( f − gγ ∧ 1)2−γ .
γ =0
Another definition of S is the following:12 Let f ∈ C ∞ (Rd ; R) and define seminorms  f  N ,n := sup (1 + r ) N (∂ n f )(r ). r ∈Rd
Then S :=
! ˆ d. f ∈ C ∞ (Rd ; R) :  f  N ,n < ∞ ∀N ∈ N, n ∈ N
One may show that the metric dS,H is equivalent to the metric
dS,sup ( f, g) := ( f − g N ,n ∧ 1)2−(N +n) , N ,n
where n = (n 1 , . . . , n d ) and n =
d
n i . Therefore, both definitions of S coincide.
i=1
Next, identify H0 with its strong dual H′0 and denote by S ′ (Rd ) the dual of S(Rd ). S ′ (Rd ) is called the Schwarz space of tempered distributions on Rd . For γ ≥ 0 we set ! H−γ := f ∈ S ′ (Rd ) : f ∈ L(Hγ ; R) 12
Cf., e.g. Treves (1967), Part II, Chap. 25.
15.1 Analysis
347
(the bounded linear functionals on Hγ ). We extend the scalar product & f, g'0 to a duality &·, ·' between S(Rd ) and S ′ (Rd ). We verify that the functional norm on H−γ , · −γ , may be defined as the norm associated with the following scalar product on H−γ :
& f, g'−γ = (d + k)−γ & f, φk '&g, φk '. ˆd k∈N
Hence, if γ1 , γ2 > 0, we obtain the chain of spaces with dense and continuous imbeddings S := S(Rd ) ⊂ Hγ1 ⊂ H0 = H′0 ⊂ H−γ2 ⊂ S ′ := S ′ (Rd ) Next, for all all γ ∈ R γ
{φk,γ } := (d + k)− 2 φk
!
(15.32)
(15.33)
is a CONS in Hγ . For applications to linear PDEs and SPDEs (cf. Sect. 13.1.1) the following characterization of the above Hilbert spaces may be useful. Let “Dom” denote the γ domain of an operator. Consider the fractional powers (−△ + r 2 ) 2 of the selfadjoint operator (−△ + r 2 )13 where γ ∈ R. For suitable restrictions or extensions γ of (−△ + r 2 ) 2 14 , denoting all restrictions and extensions with the same symbol, (15.12) and the definition of the Hilbert norms · γ imply ⎫ γ ⎪ Hγ := Dom((−△ + r 2 ) 2 ) and ⎪ ⎪ ⎪ ⎪ ⎪ : ; ⎪ γ γ ⎬ 2 2 2 2 (−△ + r  ) f, (−△ + r  ) g (15.34) 0 ⎪ ⎪
⎪ ⎪ = (d + k)γ & f, φk '&g, φk ' = & f, g'γ ∀ f, g ∈ Hγ . ⎪ ⎪ ⎪ ⎭ ˆd k∈N
In what follows assume −∞ < γ2 < γ1 < ∞. We verify that the imbedding Hγ1 ⊂ Hγ2
is compact if γ1 − γ2 > 0. A finitely additive Gauss measure on Hγ1 whose covariance operator is the identity on Hγ1 can be extended to a countably additive Gaussian measure on Hγ2 if, and only if, the imbedding is “HilbertSchmidt,” i.e., if15
φk,γ1 2γ2 < ∞. (15.35) k
13 14 15
Cf., e.g., Pazy (1983), Sect. 2.6, or Tanabe (1979), Sect. 2.3. Cf. Kotelenez (1985). Cf. Itˆo (1984), Kotelenez (1985), and Walsh (1986) in addition to Kuo (1975), Chap. I, Sect. 4, Theorem 4.3, and our previous discussion of the abstract case.
348
15 Appendix
Obviously, the imbedding is HilbertSchmidt if, and only if, γ1 − γ2 > d. The previous definitions and results imply the following: The imbedding H0 ⊂ H−γ is HilbertSchmidt, i.e., if and only if γ > d.
(15.36)
15.1.4 Metrics on Spaces of Measures Properties of various versions of the Wasserstein metric for M f and M∞,̟ as well as properties of the weight function ̟ are derived. At the end we compare the Wasserstein metric with the metric in total variation. In Chap. 4 we introduced M f as the space of finite Borel measures on B d (the Borel sets in Rd ). We define a Wasserstein metric on M f as follows: The space of all continuous Lipschitz functions f from Rd into R will be denoted C L (Rd ; R). Further, C L ,∞ (Rd ; R) is the space of all uniformly bounded Lipschitz functions f from Rd into R. Abbreviate16  f  := sup  f (q); f L := q
sup {r !=q,r −q≤1}
For µ, ν ∈ M f , we set γ f (µ − ν) :=
sup f L ,∞ ≤1
 f (r )− f (q) ρ(r −q) ;
f L ,∞ := f L ∨  f .
f (q)(µ(dq) − ν(dq)) .
(15.37)
(15.38)
Obviously (M f , γ f ) is a metric space where γ f (µ − ν) is actually a norm. Endowed with the functional norm the space of all continuous linear bounded functionals from C L ,∞ (Rd ; R) into R (denoted C L∗ ,1 (Rd ; R)) is the strong dual of C L ,∞ (Rd ; R) and, therefore, a Banach space. M f is the cone of (nonnegative) measures in C L∗ ,1 (Rd ; R), and convergence of a sequence µn ∈ M f in the normtopology implies convergence in the weak* topology. It follows from the Riesz theorem that the limit µ is a Borel (or Baire) measure. Thus, it is in M f .17 Hence, M f is closed in the normtopology and, therefore, it is also complete. Set & N
d ai δri , N ∈ N, ri ∈ R , ai ∈ R , Md := µ := i=1
i.e., Md is the space of finite sums of point measures with nonnegative weights. Obviously, Md ⊂ C L∗ ,1 (Rd ; R). 16 17
In the definition of the Lipschitz norm (15.37) we may, without loss of generality, restrict the quotient to r − q ≤ 1, since for values r − q > 1 the quotient is dominated by 2 f . Cf. Bauer (1968), Sect. VII, 45.
15.1 Analysis
349
Note that Md is dense in (M f , γ f ) in the topology generated by γ f .18 Further, we verify that the sums of point measures with rational nonnegative weights and supports in rational vectors are also dense in (M f , γ f ). Therefore, we have Proposition 15.9. (M f , γ f ) is a complete separable metric space.19 Choose ν = 0 (the Borel measure that assigns 0 to all Borel sets) in (15.38). For positive measures µ we have γ f (µ) =< 1, µ >= µ(Rd ), i.e., the norm γ f (µ) of µ ∈ M is the total mass of µ. Let us now make some comments on the relation of γ f to the Wasserstein (or MongeWasserstein) metric.20 If µ, ν have equal total finite mass m¯ > 0 we will call positive Borel measures Q on R2d joint representations of µ and ν, if Q(A × Rd ) = µ(A)m¯ and Q(Rd × B) = µ(B) ˜ m¯ for arbitrary Borel sets A, B ⊂ Rd . The set of all joint representations of (µ, µ) ˜ will be denoted by C(µ, µ). ˜ For µ, µ˜ with mass m¯ and p ≥ 1 define the pth Wasserstein metric by γ˜p (µ, µ) ˜ :=
inf
Q∈C(µ,µ) ˜
Q (dr, dq)ρ p (r − q)
1
p
.
(15.39)
For the case of two probability measures µ, ν (or finite measures of equal mass) the KantorovichRubinstein theorem asserts21 f (q)(µ(dq) − ν(dq)) = γ˜1 (µ, ν) = inf Q (dr, dq)ρ(r − q) sup f L ≤1
Q∈C(µ,ν)
(15.40) The first observation is that, both in (15.38) and (15.40), the sets of f are invariant with respect to multiplication by −1, which allows us to drop the absolute value bars in the definition of the metrics. Further, for two probability measures µ and ν sup f (q)(µ(dq) − ν(dq)). f (q)(µ(dq) − ν(dq)) ≥ sup f L ≤1
f L ,∞ ≤1
Next, for arbitrary constants c f sup ( f (q) − c f )(µ(dq) − ν(dq)) = sup f (q)(µ(dq) − ν(dq)) f L ≤1 f L ≤1 −c f (µ(Rd ) − ν(Rd )) f (q)(µ(dq) − ν(dq)), = sup f L ≤1
18 19 20 21
Cf. De Acosta, A. (1982). By Definition 15.2 and (15.38) (M f , γ f ) is a separable Fr´echet space. Cf. Dudley (1989), Chap. 11.8). Cf. Dudley (loc.cit.), Chap. 11, Theorem 11.8.2.
350
15 Appendix
since µ(Rd ) − ν(Rd ) = 0. On the lefthand side of (15.40) we may therefore, without loss of generality, assume f (0) = 0. If we choose two point measures δr , δ0 the (only) joint representation is δ(r,0) . (15.37), in addition to f (0) = 0, implies sup f L ≤1
f (r ) = ρ(r ) ≤ 1.
We obtain  f  ≤ 1
if f L ≤ 1 and if f (0) = 0.
Thus, we have for two probability measures µ and ν γ˜1 (µ, ν) =
sup f L ≤1, f (0)=0
f (q)(µ(dq)−ν(dq)) =
sup f L ,∞ ≤1
f (q)(µ(dq) − ν(dq))
and therefore, for two probability measures µ and ν γ˜1 (µ, ν) = γ f (µ − ν). As a consequence, we may in the Kantorovich–Rubinstein theorem replace the µ sup by the sup . “Normalizing” two measures µ, ν ∈ Mm¯ with m¯ > 0 to m ¯
f L ≤1 ν m¯ ,
and
f L ,∞ ≤1
respectively, we obtain γ f (µ − ν) = m¯ γ˜1
µ m¯
−
ν . m¯
(15.41)
Next, consider two arbitrary measures µ, ν ∈ M+f with total masses m¯ and n, ¯ respectively. Suppose m¯ > n¯ > 0. Then, γ f (µ − ν) = sup f (q)(µ(dq) − ν(dq)) f L ,∞ ≤1 = sup f (q)(µ(dq) − ν(dq)) f L ,∞ ≤1
(as the set of f in question is invariant if the elements are multiplied by (−1)) m¯ − n¯ n¯ µ(dq) − ν(dq) + f (q) µ(dq) = sup f (q) m¯ m¯ f L ,∞ ≤1 n¯ m¯ − n¯ µ(dq) − ν(dq) + sup µ(dq) ≤ sup f (q) f (q) m ¯ m¯ f L ,∞ ≤1 f L ,∞ ≤1 m¯ − n¯ n¯ = γf µ − ν + 1(q) µ(dq) m¯ m¯ (where 1(q) ≡ 1)
15.1 Analysis
351
n¯ µ − ν + m¯ − n¯ m ¯µ ν = n¯ γ˜1 + m¯ − n¯ − m¯ n¯
= γf
(by (15.41).) If n¯ = 0, then ν = 0. We recall that γ f (µ − ν) = γ f (µ − 0) = m. ¯ For m¯ > n¯ > 0 we now suppose mn¯¯ µ != ν which is equivalent to γ f mn¯¯ µ − ν =: n¯ c > 0. Let f n be a sequence with f n L ,∞ ≤ 1 and γ m¯ µ − ν = c = lim f n (q) mn¯¯ µ(dq) − ν(dq) . Note that
n→∞
lim
n→∞
n¯ n¯ µ(dq) − ν(dq) = lim f n (q)ν(dq) . f n (q) f n (q) µ(dq) − n→∞ m¯ m¯
Both integrals in the righthand side of the last identity are bounded sequences αn and βn , respectively. By compactness, they contain convergent subsequences, and we may assume, without loss of generality, that both αn and βn converge to real numbers α and β, respectively. Consider first the case where −β < c. This implies α > 0, whence we may assume that αn > 0 for all n. Since αn −→ α m¯ m¯ − n¯ m¯ m¯ − n¯ m¯ m¯ − n¯ f n (q) µ(dq) = αn −→ α =α − 1 > 0. m¯ n¯ m¯ n¯ m¯ n¯ Hence, n¯ n¯ µ − ν = c = lim µ(dq) − ν(dq) γf f n (q) n→∞ m¯ m¯ m¯ − n¯ n¯ ≤ lim µ(dq) − ν(dq) + lim µ(dq) f n (q) f n (q) n→∞ n→∞ m¯ m¯ = lim f (q)(µ(dq) − ν(dq)) f n (q)(µ(dq) − ν(dq)) ≤ sup n→∞
f L ,∞ ≤1
= γ f (µ − ν). Next, assume the other case −β ≥ c ⇐⇒ β ≤ −c. Since c > 0, we may assume that βn = f n (q)ν(dq) < 0 for all n. So
352
15 Appendix
n¯ n¯ µ − ν = c = lim µ(dq) − ν(dq) γf f n (q) n→∞ m¯ m¯ m¯ n¯ = f n (q) µ(dq) − ν(dq) lim m¯ n→∞ n¯ m¯ − n¯ n¯ f n (q)(µ(dq) − ν(dq)) + lim f n (q)ν(dq) = m¯ n→∞ n¯ n¯ f n (q)(µ(dq) − ν(dq)) ≤ lim m¯ n→∞ n¯ sup ≤ f (q)(µ(dq) − ν(dq)) m¯ f L ,∞ ≤1 =
n¯ γ f (µ − ν) < γ f (µ − ν). m¯
Note that for m¯ > n¯ (n¯ ∧ m)γ ¯ f
µ m¯
−
ν = γf n¯
n¯ µ−ν . m¯
The lefthand side is symmetric with respect to m¯ and n. ¯ Therefore, we obtain in both cases µ ν (n¯ ∧ m)γ ¯ f ≤ γ f (µ − ν). (15.42) − m¯ n¯ Further, note that m¯ − n¯ = 1(q)(µ(dq)−ν(dq)) ≤ sup f (q)(µ(dq)−ν(dq)) = γ f (µ−ν) f L ,∞ ≤1
Altogether,
Set
γ f (µ − ν) ≤ (n¯ ∧ m)γ ¯ f
µ
γ f (µ − ν) ≤ (n¯ ∧ m) ¯ γ˜1
µ
m¯
−
⎫ ν + m¯ − n ¯ ≤ 2γ f (µ − ν) ⎪ ⎪ ⎪ n¯ ⎪ ⎬
⇐⇒ ⎪ ⎪ ⎪ ν ⎪ + m¯ − n ¯ ≤ 2γ f (µ − ν). ⎭ − m¯ n¯ µ
(15.43)
ν + m¯ − n ¯ (15.44) m¯ n¯ We see that γˆ f (µ, ν) defines a metric on M f . An equivalent version was introduced in Kotelenez (1996). Summarizing the previous results we have shown: γˆ f (µ, ν) := (n¯ ∧ m) ¯ γ˜1
−
Proposition 15.10. γ f (µ − ν) and γˆ f (µ, ν) are equivalent metrics on M f .
⊔ ⊓
15.1 Analysis
353
Recall that M∞,̟ was defined in (4.5) as the space of all σ finite Borel measures µ on Rd such that 3 M∞,̟ := µ ∈ M∞ : ̟ (q)µ(dq) < ∞ where ̟ (r ) = (1 + r 2 )−γ , γ > d2 , is the weight function from (4.4). ̟µ etc. denotes the finite measure that can be represented as µ with density ̟ . Further, defining for µ, ν ∈ M∞,̟ γ̟ (µ − ν) := γ f (̟ (µ − ν)),
(15.45)
we obtain that (M∞,̟ , γ̟ ) is isometrically isomorphic to (M f , γ f ). Recalling Proposition 15.9, we conclude that (M∞,̟ , γ̟ ) is also a complete separable metric space. Next, we derive some properties of ̟ . For k, ℓ = 1, . . . , d and f twice continuously differentiable we recall the notation ∂k f :=
∂ ∂2 2 f, ∂kℓ f := f. ∂rk ∂rk ∂rℓ
Let β > 0 and consider ̟β (r ) := ̟ β (r ). We easily verify
⎫ ⎪ ≤ c̟β (q), ⎬ ∂k ̟β (q) + k,ℓ=1 k=1 ⎪ ⎭ ̟β (r ) − ̟β (q) ≤ γβ̺(r, q)[̟β (r ) + ̟β (q)]. d
d
2 ∂kℓ ̟β (q)
(15.46)
Indeed, a simple calculation yields
∂k ̟ (r ) = −γ (1 + r 2 )−γ −1 2rk
2 ̟ (r ) = −γ (−γ − 1)(1 + r 2 )−γ −2 2r 2r − γ (1 + r 2 )−γ −1 2δ . ∂k,ℓ k ℓ k,ℓ
This implies the first inequality. The second inequality is obvious if r − q ≥ 1. For the case r − q < 1 we may assume r  < q and β = 1. Then ̟ (r ) − ̟ (q) = ̟ (r ) − ̟ (q) q = 2γ u(1 + u 2 )−(γ +1) du r  q ≤ γ (1 + u 2 )−γ du r 
(as 2u ≤ (1 + u 2 )) q ≤ γ (1 + r 2 )−γ du r 
(as (1 + u 2 ) is monotone decreasing for u ≥ 0 )
= (q − r )̟ (r ) ≤ r − q̟ (r ) ≤ r − q[̟ (r ) + ̟ (q)].
354
15 Appendix
The following was used in (11.9): 2 γ ̟ −1 (q)̟ (r ) = 1+q ≤ 2γ (1 + r − q2 )γ . 2 1+r 
(15.47)
(15.47) follows from
̟ −1 (q) = 1 + q2 = 1 + q − r + r 2 ≤ 1 + 2q − r 2 + 2r 2 ≤ 2(1 + q − r 2 + r 2 ) ≤ 2(1 + r − q2 )(1 + r 2 ).
We have the continuous inclusion (M f , γ f ) ⊂ (M∞,̟ , γ̟ ). In what follows we compare the topologies of (M f , γ̟ ) and of (M∞,̟ , γ̟ ), respectively, with the topologies of Wm,2,Φ . As a preliminary, we derive a relation between weak and strong convergence in separable Hilbert spaces, using the Fourier expansion of the square of the Hilbert space norm.22 Lemma 15.11. Let H be a separable Hilbert space with scalar product &·, ·'H and associated Hilbert space norm ·H and let {ϕk }k∈N be a CONS in H. Let f n , f ∈ H such that (i) f n H −→ f H , as n −→ ∞, (ii) & f n , ϕk 'H −→ & f, ϕk 'H ∀k ∈ N, as n −→ ∞. Then, f n − f H −→ 0, as n −→ ∞. Proof. & f n , f 'H =
k∈b f N
& f n , ϕk 'H &ϕk , f 'H .
By assumption, the kth term in the above Fourier series converges to & f, ϕk 'H &ϕk , f 'H ∀k. To conclude that the the lefthand side also converges to & f, f 'H , we must show that the terms are uniformly integrable with respect to the counting measure. Note that & f n , ϕk 'H &ϕk , f 'H  ≤
1 [& f n , ϕk 'H &ϕk , f n 'H + & f, ϕk 'H &ϕk , f 'H ]. 2
Again our assumption implies that the first term & f n , ϕk 'H &ϕk , f n 'H converges for all k to & f, ϕk 'H &ϕk , f 'H and its sum, being equal to f n 2H , converges by assumption to f 2H . Consequently,
1 k∈N
22
2
[& f n , ϕk 'H &ϕk , f n 'H + & f, ϕk 'H &ϕk , f 'H ] −→ f 2H , as n −→ ∞ .
The result of Lemma 15.11 is actually well known, except that we replaced weak convergence by convergence of the Fourier coefficients (cf. Yosida (1968), Sect. V.1, Theorem 8).
15.1 Analysis
355
If follows that the terms 21 [& f n , ϕk 'H &ϕk , f n 'H + & f, ϕk 'H &ϕk , f 'H ] are uniformly integrable with respect to the counting measure (cf. Bauer (1968), Sect. 20, Korollar 20.5). Since these terms dominate & f n , ϕk 'H &ϕk , f 'H  we have that the terms & f n , ϕk 'H &ϕk , f 'H  are also uniformly integrable with respect to the counting measure. Hence, we may apply Lebesgue’s dominated convergence theorem and obtain
& f n , ϕk 'H &ϕk , f 'H − & f, ϕk 'H &ϕk , f 'H  −→ 0, as n −→ ∞ . k∈N
This implies & f n , f 'H −→ & f, f 'H , as n −→ ∞ ,
and similarly for & f, f n 'H . Now the statement follows as in Yosida (loc.cit.) from f n − f 2H = f n 2H + f 2H − & f n , f 'H − & f, f n 'H −→ 0, as n −→ ∞ .
⊔ ⊓
Proposition 15.12. Let α ≥ 0, (M, γ ) ∈ {(M f , γ f ), (M̟ , γ̟ )}, and let {ϕk }k∈N ⊂ S be a CONS in Hα,# , # ∈ {1, ̟ }. Suppose f n , f ∈ Hα,# ∩ M such that (i) f n α,# −→ f α,# , as n −→ ∞, (ii) γ ( f n − f ) −→ 0, as n −→ ∞. Then, f n − f α,# −→ 0, as n −→ ∞. Proof. Recalling the definitions of γ f from (15.38) and γ̟ from (15.45), respectively, we note that for all k we find a suitable constant ck such that ck ϕk L ,∞ = 1 (cf. also the proof of Theorem 8.5). Hence, convergence of f n to f in (M, γ ) implies & f n − f, ϕk 'α,# −→ 0 ∀k ∈ N, as n −→ ∞. Employing Lemma 15.11 finishes the proof.
⊔ ⊓
For our purposes we need to mention two other metrics. The first is the Prohorov metric, d p,B . This is a metric on probability measures µ, ν, defined on the Borel sets of some separable metric space B. The Prohorov metric will be described in Sect. 15.2., (15.37). The second metric is the “metric of total variation” on M f , defined as follows: µ − ν f := sup &µ − ν, 1 A ' (15.48) d {A∈B }
Apparently, · f is a norm (on the finite signed measures), restricted to M f . Proposition 15.13. γ f (µ − ν) ≤ µ − ν f ∀µ, ν ∈ M f . Proof. (i) Since M f are finite Borel measures on B d they are regular. The Riesz representation theorem23 implies that 23
Cf. Bauer (1968), Sects. 40, 41, and Folland (1984), Sect. 7.3.
356
15 Appendix
µ − ν f =
sup {ϕ≤1,ϕ∈C0 (Rd ;R)}
&µ − ν, ϕ'.
(15.49)
Further, the regularity implies that for every δ > 0 there exists a closed ball B L δ (0) with finite radius L δ and center at the origin such that µ Rd \ B L δ (0) + ν Rd \ B L δ (0) ≤ δ. Hence,
γ f (µ − ν) ≤
sup {ϕ L ,∞ ≤1}
and µ − ν f ≥ (ii) Set
ϕ(r )1 B (0) (r )(µ − ν)(dr ) + δ Lδ
sup {ϕ≤1,ϕ∈C0 (Rd ;R)}
ϕ(r )1 B (0) (r )(µ − ν)(dr ) − δ. Lδ
⎧ 1, ⎪ ⎪ ⎨ η(x) ˜ := 1 + δ(L δ − x), ⎪ ⎪ ⎩ 0,
and
if
if
0 ≤ x ≤ Lδ , x ∈ L δ + 1δ ,
otherwise,
η(r ) := η(r ˜ ).
If ϕ L ,∞ ≤ 1 we verify that the pointwise product of the two functions, ϕ and η, satisfies the following relations: ϕη ∈ C0 (Rd ; R) and ϕη L ,∞ ≤ (1 + δ). (iii) From (ii) sup ϕ(r )1 BL δ (0) (r )(µ − ν)(dr ) {ϕ L ,∞ ≤1} = sup ϕ(r )η(r )1 BL δ (0) (r )(µ − ν)(dr ) {ϕ L ,∞ ≤1} ϕ(r )η(r )1 BL δ (0) (r )(µ − ν)(dr ) ≤ sup {ϕη L ,∞ ≤1+δ,ϕ∈C L ,∞ (Rd ;R)} ϕ(r )1 BL δ (0) (r )(µ − ν)(dr ) ≤ sup {ϕ≤1+δ,ϕ∈C0 (Rd ;R)} = sup ϕ(r )1 BL δ (0) (r )(µ − ν)(dr ) (1 + δ). {ϕ≤1,ϕ∈C0 (Rd ;R)}
Therefore, from (i) we obtain
γ f (µ − ν) ≤ µ − ν f (1 + δ) + δ(2 + δ). Since δ > 0 was arbitrary the proof is complete.
⊔ ⊓
15.1 Analysis
357
15.1.5 Riemann Stieltjes Integrals Next, we provide some facts about RiemannStieltjes integrals, following Natanson (1974). Definition 15.14. Let f and g4 be realvalued functions, defined on a one5 dimensional interval [a, b] and a = t0n ≤ t1n ≤ . . . . ≤ tnn = b be a sequence of n −→ 0, as n −→ ∞. partitions such that max1≤k≤n tkn − tk−1 (I) Suppose that Vn ( f ) :=
n f (t n ) − f (t n ) converges, as n −→ ∞. Then, f k k−1
k=1
is called to be of “bounded variation” and Vab ( f ) := lim
n f (t n ) − f (t n k k−1
n→∞ k=1
the variation of f on [a, b]. n 0 / 0 /n n ) converges, (II) Let ξkn ∈ tk−1 g ξkn f (tkn ) − f (tk−1 , tkn arbitrary. Suppose k=1
as n −→ ∞. Then g is said to be “RiemannStieltjes integrable with respect to f ” and
a
b
g(t) f (dt) := lim
n→∞
n
k=1
/ 0 n g(ξkn ) f (tkn ) − f (tk−1 )
the “Riemann–Stieltjes integral” on [a, b] of g with respect to f .
(15.50) ⊔ ⊓
The following result is well known. The proof is the same as for the classical Riemann integral.24 Theorem 15.15. If g is continuous on [a, b] and f is of bounded variation on [a, b], then g is Riemann–Stieltjes integrable with respect to f and b g(t) f (dt) ≤ max g(t)Vab ( f ) ⊔ ⊓ a≤t≤b
a
The Riemann–Stieltjes integral may be employed to define the Wiener integral (cf. (15.107)) with smooth integrands. This application is based on the following integrationbyparts formula:
Theorem 15.16. g is Riemann–Stieltjes integrable with respect to f if, and only if, f is Riemann–Stieltjes integrable with respect to g and the following integrationbyparts formula holds: b b (15.51) g(t) f (dt) + f (t)g(dt) = f (b)g(b) − f (a)g(a). a
24
a
Cf.,e.g., Natanson, (loc.cit.), Chap. 8.7, Theorem 1.
358
15 Appendix
Proof. 25 Without 4loss of generality, suppose g is 5Riemann–Stieltjes integrable with n n n respect to f . Let a =n t0 ≤ t1 ≤ . . . . ≤ tn = b ben a sequence /n 0of partitions such −→ 0, as n −→ ∞, let ξk ∈ tk−1 , tkn arbitrary. Set that max1≤k≤n tkn − tk−1 Sn :=
n
/ n 0 f ξkn g tkn − g tk−1 . k=1
Obviously, Sn = whence Sn = −
n n
n
, f ξkn g tk−1 f ξkn g tkn − k=1
k=1
n−1
/ n 0 g tkn f ξk+1 − f ξkn + f ξkn g(b) − f ξ1n g(a). k=1
Adding and subtracting the righthand side of (15.51), we obtain &n−1
/ 0 / n n 0 n 0 / n n Sn = − g tk f ξk+1 − f ξk + f (b) − f ξk g(b) − f ξ1 − f (a) g(a) k=1
+ f (b)g(b) − f (a)g(a).
By assumption n−1
0 0 / / n − f ξkn + f (b) − f ξkn g(b) g tkn f ξk+1 k=1
0 / − f ξ1n − f (a) g(a) −→
a
b
f (t)g(dt), as n −→ ∞
4 5 n since a =: ξ0n ≤ ξ1n ≤ . . . . ≤ ξnn ≤ ξn+1 = b is a sequence of partitions such that n 0 / n n max1≤k≤n+1 ξk − ξk−1 −→ 0, as n −→ ∞ and tkn ∈ ξkn , ξk+1 ∀k, n. ⊔ ⊓
Observe that the Riemann–Stieltjes integral is constructed by partitioning the time axis into small intervals and then passing to the limit. This construction may be achieved in two steps. Fixing the integrator g(·), we first define the Riemann– Stieltjes integral for step function integrands f (·).26 In the second step we may or may not follow the classical procedure outlined above. An alternative method is to approximate other possible integrands “suitably” by step function integrands and extend the integral by continuity (with respect to some norm or metric) to a larger 25 26
The proof is adopted from Natanson (loc.cit.), Chap. 8.6. As we will see in Sect. 15.2.5, the stochastic Itˆo integral necessarily starts with step function type processes as elementary integrands because the integrator increments in time have to be “orthogonal” to the integrator.
15.1 Analysis
359
class of integrands. This begs the question of what is “suitable,” how large is the class of possible integrands, and what are the properties of the integral. If we wish to have a dominated convergence theorem for the resulting integral, we may first interpret the Riemann–Stieltjes integral for step functions as the Lebesgue–Stieltjes integral with respect to the (signed) Stieltjes measure associated with the integrator g(·). Representing g as the difference of two nondecreasing functions, it suffices to consider only integrators that are nondecreasing. In this case the integrator g(·) must be right continuous to entail the continuity from above of the associated Stieltjes measure and, therefore, the dominated convergence theorem. An answer to our question what is “suitable” etc. is provided by Bichteler (2002) who provides a detailed account of the extension of the Lebesgue–Stieltjes integral from step function integrands to a large class such that the dominated convergence theorem is valid. It is particularly important that, in the stochastic case, this extension leads to the stochastic Itˆo integral. In the next section we review some of the analytic properties of right continuous functions with limits from the left.
15.1.6 The Skorokhod Space D([0, ∞); B) We provide some basic definitions and properties of the Skorokhod space of cadlag functions with values in some metric space. Let (B, dB (·, ·)) be a metric space. The space C([0, ∞); B) is the natural state space for Bvalued stochastic processes with continuous sample paths. However, there are many classes of stochastic processes with jumps,27 and it is desirable to have a state space of Bvalued functions that contain both continuous functions and suitably defined functions with jumps. Skorokhod (1956) introduces such a function space and derives both analytic properties of the function space and weak convergence properties of corresponding Bvalued stochastic processes. In this section we restrict the presentation to the analytic properties of the function space. Let D([0, ∞); B) denote the space of Bvalued cadlag functions on [0, ∞) (i.e., of functions that are continuous from the right and have limits from the left).28 As for continuous functions, a suitable metric on D([0, ∞); B) will be defined on compact intervals first and then extended to a metric on [0, ∞). Therefore, let us for the moment focus on D([0, 1]; B) and C([0, 1]; B) (the space of continuous functions on [0, 1] with values in B. Remark 15.17. Let us endow C([0, 1]; B) with the uniform metric 27 28
E.g., Poisson processes or processes arising in a time discretization scheme as our Theorem 2.4 in Chap. 3. Skorokhod (1956) considers D([0, 1]; B). For more references on our presentation, cf. Kolmogorov (1956), Billingsley (1968), Chap. 3, Jacod (1985), and Ethier and Kurtz (1986), Sect. 3.5.
360
15 Appendix
du,1,B ( f, g) := sup dB ( f (t), g(t). t∈[0,1]
(15.52)
For the theory of stochastic properties the uniform metric has the following two important properties: (i) Completeness and separability of B implies completeness and separability of (C([0, 1]; B), du,1,B ). (The completeness is obvious and to see the separability it suffices to take the family of all piecewise linear continuous Bvalued functions with finitely many rational endpoints and taking values at those endpoints in a countable dense set of B.) (ii) du,1,B ( f n , f ) → 0 is equivalent to dB ( f n (tn ), f (t)) → 0 whenever tn → t. The first property is needed to obtain relative compactness criteria of families of (continuous) random processes with values in a complete and separable metric space B.29 The second property is obviously desirable for many models and their numerical approximations. Let us now (temporarily!) endow D([0, 1]; B) with the uniform metric du . (iii) We still obtain completeness of (D([0, 1]; B), du,1,B ) if B is complete.30 However, (D([0, 1]; B), du,1,B ) is not separable, even if B is separable. To verify this statement it suffices to take an arbitrary b ∈ B such that b != 0 and to consider the family { f (ξ, ·) ∈ D([0, 1]; B) : f (ξ, t) := 1[ξ,1] (t)b, ξ irrational}. If ξ1 != ξ2 then
du ( f (ξ1 , ·), f (ξ2 , ·) = dB (b, 0).
Hence, there is an uncountable set of nonintersecting nonempty open balls of radius 12 dB (b, 0), which implies the nonseparability of (D([0, 1]; B), du,1,B ). (iv) Property (ii) of (C([0, 1]; B), du,1,B ) does not hold either for (D([0, 1]; B), du,1,B ). Consider 1 > tn > t and tn ↓ t. Setting f n (t) :≡ 1[tn ,1] b and f (t) :≡ 1[t,1] b, we have du,1,B ( f n , f ) ≡ dB (b, 0), although dB ( f n (tn ), f (t)) ≡ 0 and, by assumption, tn ↓ t.
⊔ ⊓
We have convinced ourselves that the uniform metric du,1,B is not a good choice for D([0, 1]; B). On the basis of the work of Skorokhod (1956), Billingsley (1968) introduces a metric on D([0, 1]; B), yielding properties analogous to properties (i) and (ii) of (C([0, 1]; B), du,1,B ) in Remark 15.17. We will now discuss a generalization of Billingley’s metric, d D,B on D([0, ∞); B), as provided by Ethier and Kurtz (1986), Chap. 3.5, (5.1)–(5.3). Let be the set of continuous strictly increasing Lipschitz functions λ(·) from [0, ∞) into [0, ∞) such that λ(0) = 0 and 29 30
Cf. the following Theorems 15.22 and 15.23. This assertion follows from Theorem 15.19 and the fact that the uniform metric (extended to [0, ∞)) is stronger than the the metric d D,M , defined by (15.53).
15.1 Analysis
361
λ(t) − λ(s) < ∞. γ (λ) := sup log t −s 0≤s 0 and t0δ < t1δ < · · · < tk−1 δ 4 > δ. Ethier of [0, T ] such that for every T > 0 min 1≤k≤t δ ∈[0,T ]5 tkδ − tk−1 k
and Kurtz (loc.cit.), Sect. 6.3, (6.2), characterize the “equicontinuous” subsets of (D([0, ∞); B), d D,B ) in terms of the following “modulus of continuity” w ′ (a, δ, T ) := inf max tkδ
k
/sup
δ ,t δ s,t∈ tk−1 k
dM (a(s), a(t)).
(15.54)
The following theorem, proved by Ethier and Kurtz (loc.cit.), Sect. 3.6, Theorem 6.3, is a generalization of Ascoli’s theorem to (D([0, ∞); B), d D,B ) and a somewhat more general form of a theorem proved by Billingsley (1968), Chap. 3.14, Theorem 14.3: 31
Cf. Dieudonn´e (1969), Theorem 7.5.7 and, for a typical application in ODEs (the CauchyPeano existence theorem), Coddington and Levinson (1955), Chap. 1, Theorem 1.2.
362
15 Appendix
Theorem 15.20. The closure of A ⊂ D([0, ∞); B) is compact if and only if the following two conditions hold: (i) For every rational t ≥ 0, there exists a compact set K t ⊂ B such that a(t) ∈ K t for all a ∈ A. (ii) For each T > 0, lim sup w ′ (a, δ, T ) = 0. (15.55) δ→0 a∈A
⊔ ⊓
15.2 Stochastics 15.2.1 Relative Compactness and Weak Convergence We state the basic definitions and theorems regarding weak convergence and relative compactness of random variables and stochastic processes. Starting with random variables, let (, F, P) be a probability space and suppose that all our random variables are defined on (, F, P) and take values in a metric space (B, dB (·, ·)). P(B) denotes the family of Borel probability measures on B. A sequence of µn ∈ P(B) is said to “converge weakly”32 to µ ∈ P(B) if f (a)µn (da) = f (a)µ(da) ∀ f ∈ Cb (B; R) (15.56) lim n→∞ B
B
For many estimates it is useful to have a metric on P(B) such that weak convergence is equivalent to convergence in this metric. To this end, Prohorov (1956), (1.6), defines a metric on P(B) through 4 5 d p,B (µ, ν) := inf η > 0 : µ(B) ≤ ν(B η ) + η ∀B ∈ C , (15.57)
where C is the collection of closed subsets of B and 4 5 B η := a ∈ B : inf dB (a, b) < η . b∈B
(15.58)
The metric (15.57) is called the “Prohorov metric” on P(B). Prohorov (loc.cit., Theorem 1.11) proves the following
Theorem 15.21. (P(B), d p,B ) is a complete and separable metric space and convergence in (P(B), d p,B ) is equivalent to weak convergence. ⊔ ⊓ Ethier and Kurtz (loc.cit.), Chap. 3.1, Corollary 1.9, prove the following.33 32 33
The reader will notice that in functional analysis this type of convergence would be called “weak* convergence.” For a different version, employing the metric in total variation, cf. Dudley (1989), Chap. 9, Theorem 9.3.7.
15.2 Stochastics
363
Theorem 15.22. Continuous Mapping Theorem Let (Bi , dBi (·, ·)), i = 1, 2, be two separable metric spaces, and let : B1 → B2 be Borel measurable. Suppose that µn , µ ∈ P(B1 ) ∀n satisfy lim d p,B1 (µn , µ) = 0.
n→∞
Define probability measures νn , ν ∈ P(B2 ) as the images of µn and µ under the mapping : νn := µn −1 , ν := µ −1 . Let C be the set of points of B1 at which is continuous. If µ(C ) = 1, then lim d p,B2 (νn , ν) = 0.
n→∞
⊔ ⊓
A family of probability measures Q ⊂ P(B) is called “tight” if for each ǫ > 0 there is a compact set K ⊂ S such that inf µ(K ) ≥ 1 − ǫ.
µ∈Q
The following theorem was proved by Prohorov (loc.cit.), Theorem 1.12. In modern literature it is called the “Prohorov Theorem.” We adopt the formulation from Ethier and Kurtz, loc.cit., Sect. 3.2, Theorem 2.2. Theorem 15.23. Prohorov Theorem The following statements are equivalent: (i) Q ⊂ P(B) is tight. (ii) For every η > 0, there exists a compact set K ⊂ B such that, defining K η is as in (15.58), (15.59) inf µ(K η ) ≥ 1 − η. µ∈Q
(iii) Q is relatively compact, i.e., its closure in (P(M, d p,M ) is compact.
⊔ ⊓
We next provide wellknown criteria that are equivalent to weak convergence.34 Let ∂ A denote the boundary of A ⊂ B. A is said to be a Pcontinuity set if A ∈ BB and P(∂ A) = 0. Theorem 15.24. Let (B, dB ) be a separable metric space and let {Pn , P} be a family of Borel probability measures on the (B, dB ). The following conditions are equivalent: (a) Pn ⇒ P (b) lim supn→∞ Pn (F) ≤ P(F) for all closed sets F ⊂ B (c) lim infn→∞ Pn (G) ≥ P(G) for all open sets G ⊂ B (d) lim Pn (A) = P(A) for all P−continuity sets A ⊂ B n→∞
34
Cf. Ethier and Kurtz (loc.cit. Theorem 3.1) and Dudley (1989), Chap. 11, Theorem 11.3.3.
364
15 Appendix
(e) lim d p,B (Pn , P) = 0 n→∞
(f) lim γ f (Pn − P) = 0
⊔ ⊓
n→∞
Ethier and Kurtz (loc.cit.), Corollary 3.3, prove the following Corollary 15.25. Let (B, dB ) be a metric space and let (X n , Yn ), n ∈ N, and X be B × B and Bvalued random variables. If X n ⇒ X and dB (X n , Yn ) −→ 0 in ⊔ ⊓ probability, then Yn ⇒ X. We have defined weak convergence, relative compactness, etc. for Bvalued random variables in terms of their probability distributions, where all the previous assertions hold provided (B, dB ) is separable and complete. We next extend some of those definitions and results to stochastic processes. (, F, Ft , P) is called a “stochastic basis” if the set  is equipped with a σ algebra F and a probability measure P and if, in addition, there is an increasing family of σ algebras Ft , t ≤ 0 such that Fs ⊂ Ft ⊂ F, 0 ≤ s ≤ t < ∞. Ft , t ≥ 0, is called a “filtration” and we will assume that it is right continuous, i.e., that ∩s>0 Ft+s = Ft . Henceforth we assume that (, F, Ft , P), t ≥ 0, is a stochastic basis with a right continuous filtration Ft of σ algebras and that all Bvalued stochastic processes a(·) to be jointly measurable in (t, ω) and adapted to the filtration Ft , where the latter means that a(t) is Ft measurable for all t ≥ 0. Two Bvalued stochastic processes ai (·), i = 1, 2, are “Pequivalent” if P{∃t : a1 (t) != a2 (t)} = 0. This statement is equivalent to the property that these two processes are indistinguishable as elements of a metric space as follows: Let L 0,F ([0, T ] × ; B) be the space of Bvalued (adapted and jointly measurable) stochastic processes. We can endow L 0,F ([0, T ] × ; B)) with the metric d(T,);B (a1 , a2 ) :=
T 0

dB {(a1 (t, ω), a2 (t, ω)) ∧ 1}P(dω)dt.
(15.60)
Then Pequivalence of ai (·), i = 1, 2, just means that d(T,);B (a1 , a2 ) = 0 ∀T > 0. If a2 (·) is Pequivalent to a1 (·), a2 (·) is called a “modification” of a1 (·).
(15.61)
15.2 Stochastics
365
For many problems the space of jointly measurable adapted processes is unnecessarily large and, in what follows, we restrict ourselves to (D([0, ∞); B), d D,B ). By Theorem 15.19 (D([0, ∞); B), d D,B ) is a complete and separable metric space if (B, dB ) is complete and separable, and the previous results for random variables could be carried over directly to Bvalued cadlag stochastic processes. However, for stochastic processes X (·) many properties are shown in terms of the “marginals” X (t). Therefore, the next step is to derive some stochastic Ascolitype characterization in terms of relative compactness of the marginals and the modulus of continuity (15.54).35 Recall the notation from (15.58). Theorem 15.26. Stochastic Arzela–Ascoli Theorem Let {aα (·)} be a family of Bvalued processes with sample paths in D([0, ∞); B). Then {aα (·)} is relatively compact if, and only if, the following two conditions hold: (i) For every η > 0 and rational t ≥ 0, there exists a compact set K η,t ⊂ M such that 4 η 5 (15.62) inf aα (t) ∈ K η,t ≥ 1 − η. α
(ii) For every η > 0 and rational t ≥ 0, there exists a δ > 0 such that sup P{w ′ (aα , δ, T ) ≥ η} ≤ η. α
(15.63) ⊔ ⊓
The criterion (ii) of the above theorem is difficult to apply. The following criterion, equivalent to condition (ii) of Theorem 15.26, has been obtained by Kurtz (1975).36 It is especially useful when applying to families of square integrable martingales and if we may choose β = 2 as the exponent of the following (15.64).37 We denote by E[·Ft ] the conditional expectation with respect to Ft . Theorem 15.27. Let {aα (·)} be a family of Bvalued processes with sample paths in D([0, ∞); B). Then {aα (·)} is relatively compact if, and only if, condition (i) of Theorem 15.26 holds and if the following condition holds: For each T > 0, there exist β > 0 and a family {γα (δ) : 0 < δ < 1} of nonnegative random variables such that (15.64) E dM (aαβ (t + u), aα (t))Ft ≤ E γα (δ)Ft for 0 ≤ t ≤ T, 0 ≤ u ≤ δ in addition to lim sup E γα (δ) = 0. δ→0 α
35 36 37
(15.65) ⊔ ⊓
Cf. Ethier and Kurtz (loc.cit), Sect. 3.7, Theorem 7.2. Cf. also Billingsley (1968), Chap. 3.15. A proof is also found in Ethier and Kurtz (loc.cit.), Sect. 3.8, Theorem 8.6, and Remark 8.7. Cf. the next subsection for the definition of martingales.
366
15 Appendix
15.2.2 Regular and Cylindrical Hilbert SpaceValued Brownian Motions Regular and cylindrical Hilbert spacevalued Brownian motions are defined in terms of their covariance operators. The H0 valued Brownian motion, defined through correlated Brownian noise, is shown to be cylindrical. Imbedded into a weighted L 2 space it becomes a regular Brownian motion. Let H be separable Hilbert space with scalar product < ·, · >H and norm · H . A system of independent vectors {φn } ⊂ H is called a complete orthonormal system (CONS) in H if for every f ∈ H:
& f, φn 'H φn . f = n∈N
Further, a necessary and sufficient condition for completeness of {φn } is Parseval’s identity:
f 2H = & f, φn '2H ∀ f ∈ H. n∈N
A bounded, nonnegative, and symmetric linear operator Q on H is called “nu1 clear” if its nonnegative square root, Q 2 , is “HilbertSchmidt”, i.e., if
1 Q 2 φn 2H < ∞. (15.66) n∈N
As is customary, we denote by N (a, b2 ) the distribution of a realvalued Gaussian random variable with mean a and variance b2 . Definition 15.28. Brownian Motion (1) A realvalued stochastic process β1 (·) is called a Brownian motion if (i) it has a.s. continuous sample paths (ii) there is a σ > 0 such that for 0 ≤ s ≤ t β1 (t) − β1 (s) ∼ N (0, σ 2 (t − s)) and β1 (t) − β1 (s) is independent of σ (βu : 0 ≤ u ≤ s) ,
where σ (βu : 0 ≤ u ≤ s) is the σ algebra generated by all βu with 0 ≤ u ≤ s. The realvalued Brownian motion is called “standard” if σ = 1. (2) Let (i) {β1,n } be a sequence of i.i.d. Rvalued standard Brownian motions (ii) Q W a nonnegative and selfadjoint linear operator on H (iii) {φn } a complete orthonormal system (CONS) in Dom(Q W ) ⊂ H38 38
Dom(Q W ) denotes the domain of the operator Q W .
15.2 Stochastics
367
The Gaussian Rvalued random field W (·) defined by C D ∞
1 2 β1,n ϕ, Q W &W (t), ϕ'H := φn , ϕ ∈ H,
(15.67)
H
n=1
1 2 is the nonnegative selfadjoint is called an Hvalued Brownian motion, where Q W square root of Q W .39 1
2 The Hvalued Brownian motion is called “regular” if Q W is HilbertSchmidt. 1
2 Otherwise, the Hvalued Brownian motion is called “cylindrical.” If Q W = IH , the identity operator on H, the Hvalued Brownian motion is called “standard cylindrical”. Q W is called the covariance operator of the Hvalued Brownian motion. ⊔ ⊓
In what follows, we restrict ourselves to bounded selfadjoint covariance operators, which covers most infinite dimensional Brownian motions used in the SPDE literature.40 Observe that an Hvalued cylindrical Brownian motion W (·), evaluated at fixed t, does not define a countably additive measure on H. This implies that W (·) must be treated as a generalized random field.41 If a stochastic equation is driven by a cylindrical Brownian motion W (·), it is quite natural to ask whether the cylindrical Brownian motion can be imbedded into a larger (separable) Banach space B ⊃ H such that in B W (t) defines a countably additive measure. This is the approach taken by the theory of abstract Wiener spaces.42 The calculations may become easier if, ˜ such that H is instead of a Banach space, we consider a separable Hilbert space H ˜ Consider, e.g., an Hvalued cylindrical continuously and densely imbedded in H. Brownian motion with bounded covariance operator Q W . In this case, we need the ˜ to be “HilbertSchmidt,” which means imbedding H ⊂ H ∞
n=1
φn 2H˜ < ∞,
˜ 43 where {φn } is a CONS in H and the norm in the righthand side is the norm in H. ˜ can be obtained by the completion of H with respect to the norm, Such a space H generated by scalar product & f, g'H˜ :=
∞
n=1
1 & f, φn 'H &g, φn 'H . n2
The quadratic form on the lefthand side determines a symmetric positive definite 1 HilbertSchmidt operator Q 2 from H into H by setting 39 40 41 42 43
Recall that the nonnegative square root of a nonnegative selfadjoint operator is uniquely defined and selfadjoint (cf., e.g., Kato (1976), Chap. V.3.11, (3.45)). Cf. also the analysis of linear SPDEs in the space of distributions, reviewed in Chap. 13. Cf. Gel’fand and Vilenkin (1964). Cf. Gross (1965), Kuo (1975) and the references therein. Cf. Kuo (loc.cit.), Chap. 1, Sect. 4, Theorem 4.3.
368
15 Appendix 1
Q 2 φn = n1 φn ∀n or, equivalently,
:
1
1
Q 2 f, Q 2 g
;
H
:= & f, g'H˜
(15.68)
˜ W (t) defines a countably additive Gauss measure. If W (·) is the perturbation In H of some differential equation, the original differential equation must be redefined on the larger space (which may or may not be possible). An example is pointwise multiplication, which is usually well defined on function space, but not defined on the space of distributions.44 To be more precise, consider an H0 valued standard cylindrical Brownian motion W (·) which has been undoubtedly the most popular one among all cylindrical Brownian motions. The reason for this popularity is that it may be represented as t (15.69) &W (t), ϕ'0 = ϕ(r )w(dr, ds), 0
where w(dq, ds) is a realvalued Gaussian space–time white noise.45 Hence, it may serve to model space–time white noise perturbations.46 By (15.36) the H0 valued standard cylindrical Brownian motion becomes an H−γ valued regular Brownian motion if γ > d. It is well known that elements in H−γ cannot be multiplied with each other. There are other cylindrical Brownian motions of interest in applications. In fact, the perturbation by correlated Brownian motions used in this book give rise to cylindrical Brownian motions. Let us explain. Consider the kernel the Rd valued kernel G(r ), r ∈ R from the previous sections such that its onedimensional components G k are square integrable with respect to the Lebesgue measure. Define Gaussian Rvalued random fields Wk (·, ·) by t (15.70) Wk (t, r ) := G k (r − q)w(dq, ds), k = 1, . . . , d. 0
Proposition 15.29. The Gaussian random fields Wk (·, ·) from (15.70) define cylindrical H0 valued Brownian motions unless G k (r ) = 0 a.e. with respect to the Lebesgue measure. Proof. (i) Employing the series representation (4.14) for the scalar field w(dq, ds) yields &Wk (t), ϕ'0 := 44 45
46
∞
n=1
ϕ(r )
G k (r − q)φn (q)dq drβ1,n (t), ϕ ∈ H, (15.71)
Cf. Schwarz (1954). Cf. also our discussion of the bilinear SPDE in Chap. 13. Jetschke (1986) shows that the righthand side of (15.69) defines a standard cylindrical Brownian motion. Using distributional calculus, Schauml¨offel (1986) shows also that every standard cylindrical Brownian motion can be represented by the righthand side of (15.69). Cf. (4.29) and Definition 2.2. Cf. also our Chap. 13 and the SPDEs driven by cylindrical Brownian motions, which, with a few exceptions, require space dimension d = 1, unless the SPDE is linear.
15.2 Stochastics
369
Consequently, the square root of the covariance operator Q W,k is the following integral operator: 1 2 (15.72) (Q W,k φ)(r ) := G k (r − q)φ(q) dq. 1 2 For Wk (·, ·) to be regular, Q W,k must be HilbertSchmidt. In particular it must be compact (such that the squares of the eigenvalues are summable). (ii) Korotkov (1983), Chap. 2, Sect. 4, shows that such an integral operator is compact if, and only if, G k (r ) = 0 a.e. with respect to the Lebesgue measure dr . We adjust Korotkov’s proof to our notation: 1
2 Suppose Q W,k is compact. Let B ⊂ Rd be a bounded Borel set and d be its m diameter. Set r˜ := (md, 0, . . . , 0)T ∈ Rd and define functions f m (q) := (Ur˜ m 1 B )(q) := 1 B (q + r˜ m ), where m ∈ N (and Ur˜ m is the shift operator). The functions f m are bounded, integrable, and have disjoint supports. So f m −→ 0 1
2 weakly, as m −→ ∞. Since we assumed Q W,k is compact, it follows that 1
2 Q W,k f m 20 −→ 0, as m −→ ∞. Note that
2 2 21 m Q G k (r − q)1 B (q + r˜ )dq dr W,k f m = 0
=
G k (r − q)1 B (q)dq
2
2 12 1 B dr = Q W,k 0
(by the homogeneity of the kernel G k and the shift invariance of the Lebesgue measure.) 1
2 1 B 20 = 0 for all bounded Borel sets B, whence It follows that Q W,k G k (r ) = 0 a.e. ⊔ ⊓
Although the Brownian motions, defined through (15.70), are cylindrical, they are obviously not standard cylindrical. We now show that Wk (·) from (15.70) become regular Brownian motions if imbedded into a (larger) suitably weighted L 2 space such that the constants are integrable with respect to the weighted measure. A popular choice for the weight function was defined in (4.4): ̟ (r ) = (1 + r 2 )−γ , where γ > d2 .47 We obtain that ̟ (r )dr < ∞, i.e., the constants are integrable with respect to the measure ̟ (r )(r )dr . Define the separable Hilbert space ⎫ d (H0,̟ , &·, ·'0,̟ ) := (L 2 (R , ̟ (r )dr ), &·, ·'0,̟ ), ⎬ (15.73) & f, g'0,̟ := f (r )g(r )̟ (r )dr. ⎭ 47
Cf., e.g., Triebel (1978).
370
15 Appendix
Clearly, the norm · 0,̟ is weaker than the norm  · 0 and we have H0 ⊂ H0,̟
(15.74)
with dense continuous imbedding. It is therefore natural to consider the cylindrical H0 valued Brownian motions from (15.70) as H0,̟ valued. Proposition 15.30. The Brownian motions from (15.70) define regular H0,̟ valued Brownian motions. Proof. Let φ, ψ ∈ H0,̟ . The covariance operator Q W,k,̟ for Wk (t) in H0,̟ is given by ⎫ t d ⎪ ⎪ E G k (r − q)w(dq, ds)φ(r )̟ (r )dr ⎪ ⎪ ⎪ dt 0 ⎪ ⎪ t ⎪ ⎪ ⎪ ⎬ G k (r − q)w(dq, ds)ψ(r )̟ (r )dr × (15.75) 0 ⎪ ⎪ ⎪ ⎪ = G k (r − q)φ(r )̟ (r )G k (˜r − q)ψ(˜r )̟ (˜r )d˜r dr dq ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎭ =: &Q W,k,̟ φ, ψ'0,̟ .
We show that the square root of Q W,k,̟ is HilbertSchmidt. Let {φn,̟ } be a CONS in H0,̟ . By the dominated convergence and Fubini’s theorems, in addition to the shift invariance of the Lebesgue measure, ⎫ D ∞ C 1
1 ⎪ ⎪ 2 2 ⎪ Q W,k,̟ φn,̟ , Q W,k,̟ φn,̟ ⎪ ⎪ ⎪ 0,̟ ⎪ n=1 ⎪ ⎪ ∞ ⎪
⎪ ⎪ ⎪ ⎪ &Q W,k,̟ φn,̟ , φn,̟ '0,̟ = ⎪ ⎪ ⎪ ⎪ n=1 ⎪ ⎪ ∞ ⎪
⎪ ⎪ G k (r − q)φn,̟ (r )̟ (r )G k (˜r − q)φn,̟ (˜r )̟ (˜r )d˜r dr dq ⎪ = ⎪ ⎪ ⎪ ⎬ n=1 $ %
∞ = G k (r − q)φn,̟ (r )̟ (r )G k (˜r − q)φn,̟ (˜r )̟ (˜r )d˜r dr dq ⎪ ⎪ ⎪ ⎪ ⎪ n=1 ⎪ ⎪ ⎪ ⎪ 2 ⎪ ⎪ = G k (r − q)̟ (r )dr dq ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ 2 ⎪ = G k (r − q)dq ̟ (r )dr ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ 2 ⎪ ⎭ = G k 0 ̟ (r )dr < ∞.
(15.76) ⊔ ⊓
Since the larger space H0,̟ is itself a space of functions, a number of operations, like pointwise multiplication, originally defined on H0 , can be extended to operations on H0,̟ .
15.2 Stochastics
371
The procedure leading to (15.68) shows that imbeddings from one Hilbert space ˜ can always be defined through some posH into a possibly larger Hilbert space H 1 itive definite symmetric bounded operator on Q 2 from H into H. The imbedding 1 H0 ⊂ H0,̟ from (15.74) defines the multiplication operator Q 2 √ 1 (15.77) (Q 2 f )(r ) := ̟ (r ) f (r ) r ∈ Rd , f ∈ H0 . We have
∂k ̟ (r ) = −γ (1 + r 2 )−γ −1 2rk . So ∇̟ (r ) != 0 Lebesguea.e. This implies that the spectrum of the multiplica1 tion operator Q 2 from (15.77) is (absolutely) continuous (cf. Kato (1976), Chap. 10, 1 Sect. 1, Example 1.9). Therefore, it cannot be compact and, a fortiori, Q 2 cannot 1 be a HilbertSchmidt operator Q 2 from H into H. We conclude that the imbedding (15.77) is not a HilbertSchmidt imbedding. The same result holds for any (smooth) and integrable weight function λ(r ) as long as the ∇λ(r ) != 0 Lebesguea.e.
15.2.3 Martingales, Quadratic Variation, and Inequalities We collect some definitions and properties related to martingales, semimartingales, and quadratic variation, which are standard in stochastic analysis but may cause problems to the reader from a different background. We also state Levy’s characterization of a Brownian motion, the martingale central limit theorem as well as three maximal inequalities. The first two inequalities are Doob’s and the Burkholder– Davis–Gundy inequalities. The last is a maximal inequality for Hilbertvalued stochastic convolution integrals. In this section (H, · H ) is a separable real Hilbert space48 with scalar product < ·, · >H and norm · H . Definition 15.31. Martingale An Hvalued integrable stochastic process49 m(·) is called an “Hvalued mean 0 martingale” with respect to the filtration Ft , if for all 0 ≤ t < ∞ E(m(t)Fs ) = m(s) a.s. and m(0) = 0 a.s.
⊔ ⊓
An Hvalued martingale m(·) always has a cadlag modification, i.e., there is an Hvalued martingale m(·) ¯ that is Pequivalent to m(·). It is customary to work with the “smoothest” modification of a stochastic process, which, in case of an arbitrary Hvalued martingale, is the cadlag modification.50 48 49 50
H can also be finite dimensional in this setting. In particular, it can be Rd , d ≥ 1, endowed with the usual scalar product or for d = 1 with the usual distance. Integrability here means that Em(t)H < ∞ ∀t ≥ 0. Cf. Metivier and Pellaumail (1980), Sects. 1.16 and 1.17. We remind the reader of the analogy to Sobolev spaces over Rd , Hm , defined earlier. If m > d2 , the equivalence classes of Hm always contain one continuous element. We may call this a continuous modification and assume that all elements of Hm are continuous.
372
15 Appendix
Henceforth, we will always assume that our martingales are cadlag. Since for many applications (such as integration) it suffices to consider m(·) − m(0), we may also, without loss of generality, assume that our martingales are mean 0. Within the class of (cadlag) martingales, a special role is assigned to those martingales that have a continuous modification (i.e., where almost all sample paths are continuous). These martingales are called “continuous martingales.” If, for all t m(t) is square integrable, m(·) is called an “Hvalued square integrable martingale.” If there is a sequence of localizing stopping times τn −→ ∞ with probability 1 such that m(· ∧ τn ) is a square integrable martingale, m(·) is called a “locally square integrable martingale.” Further, we recall the following terminology: b(·) is an “Hvalued process of bounded variation,” if with probability 1 the sample paths b(·, ω) have bounded variation on any finite interval [0, T ]. If b(·) has a cadlag (continuous) modification, then b(·) is called an “Hvalued cadlag (continuous) process of bounded variation.” The sum of an Hvalued process of bounded variation and an Hvalued martingale is called an “Hvalued semimartingale.” It is obvious how to extend the local square integrability (or, more generally, pintegrability for p ≥ 1) to semimartingales. Probably the most important inequalities in martingale theory and stochastic Itˆo integration are the following. Theorem 15.32. Submartingale and Doob Inequalities Let m(·) be an Hvalued martingale.51 Then for any stopping time τ (i) P{ sup m(t)H > L} ≤ 0≤t≤τ
(ii)
E sup m(t)2H 0≤t≤τ
1 L
⎫ Em(τ )H (submartingale inequality); ⎪ ⎬
≤ 4Em(τ )2H
(Doob’s inequality).
⎪ ⎭
(15.78) The proof is found in Metivier and Pellaumail (loc.cit.), Sects. 4.8.2 and 4.10.4. ⊓ ⊔ Next we define processes of finite quadratic variation (cf. Metivier and Pellaumail (loc. cit.), Chap. 2.3). Definition 15.33. Let {t0n < t1n < · · · < tkn < · · ·} be a sequence of partitions of n ) −→ 0, as n −→ ∞ and [0, ∞) such that for every T > 0 max{1≤k≤T } (tkn − tk−1 n for all n tk −→ ∞, as k −→ ∞. An Hvalued process a(·) is said to be of “finite quadratic variation” if there exists a monotone increasing realvalued process [a(·)] such that for every t > 0 n a(tkn ∧ t) − a(tk−1 ∧ t)2H −→ [a(t)] in probability, as n −→ ∞ . k≥0
(15.79) ⊔ ⊓
51
A realvalued cadlag process, x(·), is called a “submartingale” if E(x(t)Fs ) ≥ x(s) a.s. for 0 ≤ s ≤ t. The sum of submartingales is a submartingale. If m(·) is an Hvalued martingale then m(·)H is a submartingale.
15.2 Stochastics
373
In what follows, let m(·) be an Hvalued square integrable martingale. Our goal is to show that m(·) has finite quadratic variation. Recall the space L 0,F (D([0, T ]; H)), introduced before Definition 4.1. We endow L 0,F (D([0, T ]; H)) with the metric of convergence in probability dprob,D,H ( f, g) := inf P{d D,H ( f, g) ≥ ε} ≤ ε. ε>0
(15.80)
Proposition 15.34. Suppose m(·) is an Hvalued square integrable martingale. Then there is a unique quadratic variation [m](·) such that (15.79) holds with a(·) := m(·). If m(·) is continuous, then its quadratic variation [m](·) is also continuous. Proof. (i) To avoid cumbersome notation describing the jumps in cadlag processes, we provide the proof only for continuous martingales and refer the reader for the more general case to Metivier and Pellaumail, loc.cit. (ii) Let {t0n < t1n < · · · < tkn < · · ·} be a sequence of partitions of [0, ∞) as in Definition 15.33. We easily verify n m(tkn ∧ t)2H − m(tk−1 ∧ t)2H n n n = m(tkn ∧ t) − m(tk−1 ∧ t)2H +2&m(tk−1 ∧ t), m(tkn ∧ t) − m(tk−1 ∧ t)'H .
Therefore, ⎫
n m(tkn ∧ t)2H − m(tk−1 ∧ t)2H ⎪ ⎪ ⎪ k ; ⎬ :
n n n ∧ t) ∧ t), m(tkn ∧ t) − m(tk−1 = ∧ t)2H + 2 m(tk−1 m(tkn ∧ t) − m(tk−1
m(t)2H =
k =: Sn,1 (t) + 2Sn,2 (t).
H⎪ ⎪
⎪ ⎭
(15.81) Hence, for fixed t, Sn,1 (t) will converge, if Sn,2 (t) does and, of course, vice versa. However, Sn,2 (t) turns out to be a continuous martingale, whence we may use orthogonality of increments, working with squares of the sequence. Therefore, we show the convergence of Sn,2 (t). (iii) Observe that for 0 ≤ s ≤ t E(&m(s), m(t)'H Fs ) = m(s)2H . a.s.
(15.82)
Indeed, with probability 1 E(&m(s), m(t)'H Fs ) = E(& f, m(t)'H Fs ){m(s)= f } (cf. Gikhman and Skorokhod (1971), Chap. 1.3) = & f, E(m(t)Fs )'H = & f, m(s)'H{m(s)= f } = m(s)2H , where the interchangeability of the scalar product and the conditional expectation follows from the linearity and continuity of the scalar product.
374
15 Appendix
(iv) Let s < t. Then E(Sn,2 (t)Fs ) = =
k
k
+
n n E(&m(tk−1 ∧ t), m(tkn ∧ t) − m(tk−1 ∧ t)'H Fs ) n n ∧ t)'H Fs ) ∧ t), m(tkn ∧ t) − m(tk−1 1{tkn ∧t≤s} E(&m(tk−1
k≥0
n n 1{tkn ∧t>s} E(&m(tk−1 ∧ t), m(tkn ∧ t) − m(tk−1 ∧ t)'H Fs )
=: E(Sn,2,1 (t)Fs ) + E(Sn,2,2 (t)Fs ). Obviously, Sn,2,1 (t) ≡ Sn,2 (s) a.s.,
which is Fs measurable. Using standard properties of the conditional expectation in addition to (15.82), we obtain with probability 1 E(Sn,2,2 (t)Fs )
n n n ∧t) }'H Fs ) 1{tkn ∧t>s} E(&m(tk−1 ∧ t)E{(m(tkn ∧ t) − m(tk−1 ∧ t))Fs∨(tk−1 = k≥0
=
k≥0
n n n ∧ t), m(s ∨ (tk−1 ∧ t)) − m(s ∨ (tk−1 ∧ t))'H Fs ) 1{tkn ∧t>s} E(&m(tk−1
= 0.
Combining the previous calculations, we have shown that for 0 ≤ s ≤ t < ∞ E(Sn,2 (t)Fs ) = Sn,2 (s) a.s.
(15.83)
As all quantities involved in the definition of Sn,2 (·) are continuous, Sn,2 (·) is continuous and, obviously, square integrable and Sn,2 (0) = 0 by construction. Thus, we have shown that Sn,2 (·) is a realvalued continuous square integrable mean zero martingale. (v) We must show that Sn,2 (·) is a Cauchy sequence in L 0,F (C([0, T ]; H)) ∀T >0 (cf. before Definition 4.1), where the metric on L 0,F (C([0, T ]; H)) is now the metric of convergence in probability & dprob,u,T,H ( f, g) := inf P ε>0
52
sup f (t) − g(t)H ≥ ε ≤ ε.
(15.84)
0≤t≤T
Analyzing the distance between Sn,2 (· ∧ τn,δ,N ) and Sm,2 (· ∧ τn,δ,N ) for two partitions {t0n < t1n < · · · < tkn < · · ·} and {t0m < t1m < · · · < tkm < · · ·}, respectively, we may, without loss of generality, assume that52 4n 5 5 4 t0 < t1n < · · · < tkn < · · · ⊃ t0m < t1m < · · · < tkm < · · · .
Otherwise we could compare both Sn,2 (·) and Sm,2 (·) with the corresponding sum, based on the union of the partitions, {t0n < t1n < · · · < tkn < · · ·} ∪ {t0m < t1m < · · · < tkm < · · ·}.
15.2 Stochastics
375
We then rewrite Sm (·) by adding the missing partition points from {t0n < t1n < · · · < tkn < · · ·} and setting t nk := maxℓ {tℓm : tℓm ≤ tkn }, ∀k.
As in the derivation of (15.83), we then see that
Sn,2 (t) − Sm,2 (t)
A B n n = m(tk−1 ∧ t) − m(t nk−1 ∧ t), m(tkn ∧ t) − m(tk−1 ∧ t) H k
is itself a realvalued continuous square integrable martingale. For δ > 0 define the following stopping times & δ n 2 n τn,δ := inf 0 ≤ s ≤ 2T : sup m(tk−1 ∧ s) − m(t k−1 ∧ s)H ≥ Em(T )2H + 1 k and for N > 0 τn,N := inf{0 ≤ s ≤ 2T : Setting
k
n m(tkn ∧ s) − m(tk−1 ∧ s)2H ≥ N }.
τn,δ,N := τn,δ ∧ τn,N , the continuity of m(·) implies, τn,δ,N −→ 2T, a.s., as n −→ ∞ .
(15.85)
Further, also Sn,2 (t ∧τn,δ,N )− Sm,2 (t ∧τn,δ,N ) is a realvalued continuous square integrable martingale. By (15.85) it suffices to show that Sn,2 (· ∧ τn,δ,N ) − Sm,2 (· ∧ τn,δ,N ) tends to 0 uniformly on compact intervals in the metric of convergence in probability. Employing Doob’s inequality (Theorem 15.32), we obtain that for arbitrary T > 0 E 0≤t≤T (Sn,2 (t ∧ τn,δ,N ) − Sm,2 (t ∧ τn,δ,N ))2 ≤ 4E(Sn,2 (T ∧ τn,δ,N ) −Sm,2 (T ∧ τn,δ,N ))2
(15.86)
Note that, as in (15.82), for k < ℓ: n E&m(tk−1 ∧ t ∧ τn,δ,N ) − m(t nk−1 ∧ t ∧ τn,δ,N ), m(tkn ∧ t ∧ τn,δ,N ) n ∧ t ∧ τn,δ,N )'H −m(tk−1 n ∧t ∧τ n n ×&m(tℓ−1 n,δ,N ) − m(t ℓ−1 ∧ t ∧ τn,δ,N ), m(tℓ ∧ t ∧ τn,δ,N ) n ∧t ∧τ −m(tℓ−1 n,δ,N )'H n ∧ t ∧ τn,δ,N ) − m(t nk−1 ∧ t ∧ τn,δ,N ), m(tkn ∧ t ∧ τn,δ,N ) = E&m(tk−1 n n ∧t ∧τ n ∧ t ∧ τn,δ,N )'H × &m(tℓ−1 −m(tk−1 n,δ,N ) − m(t ℓ−1 ∧ t ∧ τn,δ,N ), n ∧t ∧τ n ∧t∧τ )'H = 0 ×E(m(tℓn ∧ t ∧ τn,δ,N ) − m(tℓ−1 n,δ,N )Ftℓ−1 n,δ,N n ∧t ∧τ n ∧t∧τ a.s., since E(m(tℓn ∧ t ∧ τn,δ,N ) − m(tℓ−1 )=0 n,δ,N )Ftℓ−1 n,δ,N
376
15 Appendix
Hence, E(Sn,2 (T ∧ τn,δ,N ) − Sm,2 (T ∧ τn,δ,N ))2
n =E &m(tk−1 ∧ T ∧ τn,δ,N ) − m(t nk−1 ∧ T ∧ τn,δ,N ), m(tkn ∧ T ∧ τn,δ,N ) k
=
n −m(tk−1
∧ T ∧ τn,δ,N )'H
32
n E&m(tk−1 ∧ T ∧ τn,δ,N ) − m(t nk−1 ∧ T ∧ τn,δ,N ), m(tkn ∧ T ∧ τn,δ,N ) k n n ∧ T ∧ τn,δ,N )'2H + E&m(tk−1 ∧ T ∧ τn,δ,N ) −m(tk−1 k!=ℓ
n −m(t nk−1 ∧ T ∧ τn,δ,N ), m(tkn ∧ T ∧ τn,δ,N ) − m(tk−1 ∧ T ∧ τn,δ,N )'H n ∧T ∧τ n n ×&m(tℓ−1 n,δ,N ) − m(t ℓ−1 ∧ T ∧ τn,δ,N ), m(tℓ ∧ T ∧ τn,δ,N )
n ∧T ∧τ −m(tℓ−1 n,δ,N )'H
n = ∧ T ∧ τn,δ,N ) − m(t nk−1 ∧ T ∧ τn,δ,N ), m(tkn ∧ T ∧ τn,δ,N ) E&m(tk−1 k
n −m(tk−1 ∧ T ∧ τn,δ,N )'2H
(since the mixed terms integrate to 0, as shown in the previous formula)
n Em(tk−1 ∧ T ∧ τn,δ,N ) − m(t nk−1 ∧ T ∧ τn,δ,N )2H m(tkn ∧ T ∧ τn,δ,N ) ≤ k
n ∧ T ∧ τn,δ,N )2H −m(tk−1
(by the CauchySchwarz inequality) ≤
k
n Em(tkn ∧ T ∧ τn,δ,N ) − m(tk−1 ∧ T ∧ τn,δ,N )2H
by the definition of τn,δ,N .
δ Em(T )2H + 1
By (15.81) and (15.82) n Em(tkn ∧T ∧τn,δ,N )−m(tk−1 ∧T ∧τn,δ,N )2H = Em(T ∧τn,δ,N )2H ≤ Em(T )2H .
Altogether, E(Sn,2 (T ∧ τn,δ,N ) − Sm,2 (T ∧ τn,δ,N ))2 ≤ δ.
(15.87)
Employing the Chebyshev inequality in addition to (15.86), we have shown that Sn,2 (·) is a Cauchy sequence in the metric (15.84), and it follows that the limit, S2 (·), is also a continuous realvalued square integrable martingale. We conclude ⊔ ⊓ that Sn,1 (·) converges uniformly in L 0,F (C([0, T ]; R) for all T > 0 as well. Returning to the more general case of cadlag martingales we denote the limit of Sn,2 (·) by
15.2 Stochastics
377
S2 (t) =:
t 0
&m(s−), m(ds)'H ,
(15.88)
where m(s−) is the limit from the left of m(s) (cf. Metivier and Pellaumail (loc. cit.)). The integral on the righthand side of (15.88) is called the “stochastic Itˆo integral” of the linear operator &m(s−), ·'H with respect to m(ds) (cf. also the following Sect. 15.2.5 for more details). Summarizing, it follows that t (15.89) m2H (t) ≡ 2 &m(s−), m(ds)'H + [m](t). 0
Remark 15.35. We emphasize that, in general, the quadratic variation is not trivial, i.e., it is not identically 0. (Cf. the following example.) However, we easily see that for a continuous process of bounded variation b(·) its quadratic variation must be identically 0 a.s. On the other hand, comparing the definitions of the variation of a function and the quadratic variation of processes, it is an easy exercise to verify the following statement: If the square integrable martingale m(·) is continuous and [m](·, ω) is nontrivial, then m(·, ω) is of unbounded variation a.s. ⊔ ⊓ Example 15.36. Let β(·) be an Rd valued Brownian motion with covariance matrix σ 2 Id (cf. Definition 15.28, applied to the finite dimensional Hilbert space Rd ), i.e., for 0 ≤ s ≤ t β(t) − β(s) ∼ N (0, σ 2 (t − s)Id ). Further, the increments β(t) − β(s) are independent of β(u) for 0 ≤ u ≤ s ≤ t. Therefore, β(·) is an Rd valued continuous martingale. For fixed t we can compute its quadratic variation directly, employing the strong law of large numbers (LLN) for i.i.d. random variables. With the notation of Definition 15.33 the law of large numbers (LLN) implies 2 n β(t ∧ t) − β(t n ∧ t) − tdσ 2 −→ 0, as n −→ ∞ , (15.90) k−1 k k≥0
whence,
[β](t, ω) ≡ tdσ 2 a.s.
(15.91)
[a1 , a2 ](·) := 14 [a1 + a2 ](·) − 14 [a1 − a2 ](·).
(15.92)
Observe that, by (15.79), the quadratic variation defines a bilinear form as follows: Let a1 (·) and a2 (·) be two Hvalued processes with finite quadratic variations [a1 ](·) and [a2 ](·), respectively. We easily verify that the existence of [a1 ](·) and [a2 ](·) implies the existence of both [a1 + a2 ](·) and [a1 − a2 ](·). Consequently, the following bilinear form is well defined:
The lefthand side of (15.92) is called the “mutual quadratic variation of a1 (·) and a2 (·). If b(·) is an Hvalued process of bounded variation and m(·) is an Hvalued square integrable martingale [b, m](·) ≡ 0 a.s.
(15.93)
Expression (15.93) follows in the continuous case relatively easily from the preceding calculations. We refer to Metivier and Pellaumail (loc.cit.), Sect. 2.4.2, for the general case.
378
15 Appendix
We call two Hvalued square integrable martingales, m 1 (·) and m 2 (·) “uncorrelated” if (15.94) E&m 1 (t), m 2 (t)'H = 0 ∀t ≥ 0. For uncorrelated Hvalued square integrable martingales m i (·) i = 1, 2, [m 1 , m 2 ](·) ≡ 0 a.s.
(15.95)
It follows immediately that (15.95) holds if m i (·) are independent. If m i (·) are two Hvalued square integrable martingales and ϕi are two elements of H∗ , the strong dual of H,53 we verify that (m i (·), ϕi ) are realvalued square integrable martingales, where i = 1, 2, and (·, ·) is the duality which extends the scalar product &·, ·'H . Hence, following Metivier and Pellaumail (loc.cit.), Sec. 2.3.6 and 2.4.4, we define the “mutual tensor quadratic variation” of the two martingales m i (·) as a random bilinear functional, acting on the tensor product H∗ ⊗ H∗ , by [m 1 , m 2 ](ϕ1 ⊗ ϕ2 )(·) := [(m 1 , ϕ1 ), (m 2 , ϕ2 )](·).
(15.96)
By (15.95) [m 1 , m 2 ](·) ≡ 0 a.s. if m i (·) i = 1, 2,
(15.97)
are uncorrelated Hvalued square integrable martingales. Set for the tensor quadratic variation of m(·) [m](·) := [m, m](·).
Take a CONS {ϕn } in H∗ . The trace of the bilinear functional [m](t) is defined as ∞
[(m(t), ϕn ), (m(t), ϕn )], Trace([m](t) = n=1
whence,
Trace([m](·) = [m](·) a.s.
(15.98)
The following characterization of Brownian motion is a special case of Theorem 1.1 of Ethier and Kurtz (loc.cit.), Chap. 7.1, where the covariance matrix may be time dependent. Theorem 15.37. Levy’s characterization of Brownian motion (i) Suppose m(·) is an Rd valued continuous square integrable martingale and there is a nonnegative definite symmetric matrix C ∈ Md×d such that the tensor quadratic variation satisfies [m](t) ≡ tC a.s.
(15.99)
Then m(·) is an Rd valued Brownian motion such that for all t > 0 m(t) ∼ N (0, tC). 53
Most of the time we may choose ϕi ∈ H since H and H∗ are isometrically isomorphic. However, working with chains of Hilbert spaces as in (15.32) a natural choice could be H := H−γ for some γ > 0. In this case H∗ = Hγ .
15.2 Stochastics
379
The proof uses the Itˆo formula and will be given at the end of Sect. 15.6.4.
⊔ ⊓
Remark 15.38. We remark that an Hvalued Brownian motion W (·) is Hvalued square integrable martingale if, and only if, it is regular. In this case the tensor quadratic variation is given by [W ](t) ≡ t Q W a.s.
(15.100) ⊔ ⊓
A somewhat more general version of the following martingale central limit theorem has been proved by Ethier and Kurtz (loc.cit.), Chap. 7.1, Theorem 1.4. Theorem 15.39. Martingale Central Limit Theorem (i) Let m(·) be a continuous Rd valued Gaussian meanzero martingale with deterministic continuous tensor quadratic variation [m](·) and components [m]kℓ (·), k, ℓ = 1, . . . , d. (ii) For n = 1, 2, . . . let {Ftn } be a filtration and m n (·) a locally square integrable Ftn martingale with sample paths in D([0, ∞); Rd ) and m n (0). Let [m n ](·) denote the tensor quadratic variation of m n (·) with components [m n ]kℓ (·), k, ℓ = 1, . . . , d. Suppose that the following three conditions hold: For each T > 0 and k, ℓ = 1, . . . , d / 0 lim E sup [m n ]kℓ (t) − [m]kℓ (t−) = 0. n→∞
t≤T
(15.101)
/ 0 lim E sup m n (t) − m n (t−) = 0.
(15.102)
[m n ]kℓ (t) −→ ckℓ (t) in probability.
(15.103)
m n (·) ⇒ m(·) in D([0, ∞); Rd ).
(15.104) ⊔ ⊓
n→∞
t≤T
For each t ≥ 0 and k, ℓ = 1, . . . , d
Then
We conclude this section with two additional maximal inequalities. For realvalued locally square integrable martingales we have the following improvement of Doob’s inequality: Theorem 15.40. BurkholderDavisGundy Inequality There exist universal positive constants c p , C p , where 0 < p < ∞ such that for every realvalued locally integrable martingale m(·) and stopping time τ ≥ 0 c p E sup m2 p (s) ≤ E[m] p (τ ) ≤ C p E sup m2 p (s). 0≤s≤τ
0≤s≤τ
(15.105)
380
15 Appendix
The proof for continuous locally square integrable martingales is found in Ikeda and Watanabe (1981), Chap. III.3, Theorem 3.1. the proof of the more general version of just locally integrable martingales is provided by Liptser and Shiryayev (1986) , Chap. 1.9, Theorems 6 and 7.54 ⊔ ⊓ We present a special case of an inequality proved by Kotelenez (1984), Theorem 2.1. For its formulation we require the definition of a strongly continuous (twoparameter) semigroup on a real separable Hilbert space with scalar product &·, ·'H and norm · H . Let IH denote the identity operator on H. Definition 15.41. A family of bounded linear operators on H, U (t, s) with 0 ≤ s ≤ t < ∞, is called a “strongly continuous twoparameter semigroup” on H if the following conditions hold: (i) U (t, t) ≡ IH (ii) U (t, u)U (u, s) = U (t, s) for 0 ≤ s ≤ u ≤ t < ∞ (iii) U (t, s) is strongly continuous in both s and t 55 The uniform boundedness principle (Theorem 15.3, Part (II)) implies sup U (t, s)L(H) < ∞ ∀t¯ < ∞.
0≤s≤t≤t¯
⊔ ⊓
Theorem 15.42. Maximal Inequality for Stochastic Convolution Integrals Let m(·) be an Hvalued square integrable cadlag martingale and U (t, s) a strongly continuous twoparameter semigroup of bounded linear operators. Suppose there is an η ≥ 0 such that
U (t, s)L(H) ≤ eη(t−s) ∀ 0 ≤ t < ∞. · Then the Hvalued convolution integral 0 U (·, s)m(ds) has a cadlag modification and for any bounded stopping time τ ≤ T < ∞ t U (t, s)m(ds)2H ≤ e4T η E[m](τ ). E sup (15.106) 0≤t≤τ
0
⊔ ⊓
15.2.4 Random Covariance and Space–time Correlations for Correlated Brownian Motions The random covariance of both generalized and classical random processes and random fields, related to Rd valued correlated Brownian motions, are analyzed. For 54
Both Liptser and Shiryayev (loc.cit.) and Metivier and Pellaumail (loc.cit.) prove additional inequalities for martingales. 55 The definition is a straightforward generalization of strongly continuous oneparameter semigroups. For the definition and properties of strongly continuous oneparameter and twoparameter groups we refer to Kato (1976) and Curtain and Pritchard (1978). However, we do not make assumptions about the existence of a generator A(t). Curtain and Pritchard (1978) call U (t, s) called a “mild evolution operator.”
15.2 Stochastics
381
the time derivative of correlated and uncorrelated Brownian motions the Wiener integral is employed. We compare the results with the covariance of uncorrelated Brownian motions. We start with the time correlations of a family of square integrable Rd valued continuous martingales martingales m 1 (·), m 2 (·), which are adapted to the same filtration Ft . Note that the mutual tensor quadratic variation [m i , m j ], i, j = 1, 2, is absolutely continuous and, therefore, differentiable in the generalized sense. Similar to the case of an R2 valued Brownian motion (β1 (·), β2 (·)), we treat the time derivative of m i (·) as a random distribution over R (also called “generalized random processes”) and follow the analysis provided by Gel’fand and Vilenkin56 for a onedimensional Brownian motion. The main difference from the case of (β1 (·), β2 (·)) is that m 1 (·), m 2 (·) is not necessarily Gaussian. Therefore, instead of computing the covariance directly, we may in a first step work with the mutual quadratic variation and then determine the covariance, taking the mathematical expectation of the mutual quadratic variation. Let η, ϕ, and ψ be test functions from Cc∞ (R; R), i.e., from the space infinitely often differentiable realvalued functions on R with compact support. Denoting by &ϕ, F' the duality between a distribution and a test function, the duality between the random distribution (or generalized random process) d i m (·) and the test function η is given by the Wiener integral, i.e., dt C D ∞ d η, m i (·) := η(t)dm i (t), i = 1, 2. (15.107) dt 0 Integration by parts, in the sense of Riemann–Stieltjes integrals,57 yields T T d i η(t) m ik (t)dt +η(T )m ik (T ), η(t)dm k (t) = − dt 0 0 k = 1, . . . , d, i = 1, 2, ∀T > 0, where we used m ik (0) = 0 to simplify the boundary condition.58 Letting T −→ ∞, the fact that η has finite support implies: ∞ ∞ d i η(t)dm k (t) = − ( η(t))m ik (t)dt, k = 1, . . . , d, i = 1, 2. (15.108) dt 0 0
Since the test functions have finite support we obtain the existence of the limit of T T j the mutual quadratic variation [ 0 ϕ(t)dm ik (t), 0 ψ(t)dm ℓ (t)], as T −→ ∞.
Definition 15.43. Cov ω
∞ d d ϕ, m i ψ, m j := ϕ(t)dm ik (t), dt dt 0 kℓ ∞ j ψ(t)dm ℓ (t) i, j = 1, 2, k, ℓ = 1, . . . , d,
(15.109)
0
56 57 58
Cf. Gel’fand and Vilenkin (1964), Chap. III.2.5. Cf. our Theorem 15.16. The Riemann–Stieltjes interpretation remains obviously valid, if we just assume η(·) to be continuous and of bounded variation on [0, T ].
382
15 Appendix
will be called the “random covariance” of the generalized random processes the righthand side of (15.109) is, by definition, equal to j ψ(t)dm ℓ (t)]
d i i = 1, 2, where dt m (·), T T lim [ 0 ϕ(t)dm ik (t), 0 T →∞
Generalizing Gel’fand and Vilenkin (loc.cit), we compare the random covariance d 2 d 1 m (·), m (·) with random covariance of (m 1 (·), m 2 (·)), where of dt dt d d ϕ, m i ψ, m j Cov ω dt dt kℓ ∞ ∞ d d j := ϕ (t) ψ (s)[m ik , m ℓ ](t ∧ s)dt ds i, j = 1, 2, dt ds 0 0 (15.110) By the orthogonality of the martingale increments the righthand side of (15.110) equals ∞ ∞ d d j ϕ (t) ψ (s)[m ik , m ℓ ](t ∧ s)dt ds dt ds 0 0 ∞ ∞ d d j i ϕ (t)m k (t)dt, ψ (s)m ℓ (s)ds . (15.111) = dt dt 0 0 Theorem 15.44. ⎫ d d d d ⎪ ψ, m j = Cov ω ϕ, m i ψ, m j ⎪ Cov ω ϕ, m i ⎬ dt dt dt dt ∞ ⎪ d ⎭ = ϕ(t)ψ(t) [m i , m j ](t)dt, i, j = 1, 2. ⎪ dt 0 (15.112) Proof. The first equality in (15.112) follows from integration by parts in the generalized sense. To show the second equality we set f (t) := [m i , m j ](t)
and follow the procedure of Gel’fand and Vilenkin (loc.cit) who consider for Brownian motion the case f (t) ≡ t. We obtain ∞ ∞ d d ϕ (t) ψ (s) f (t ∧ s)dt dt dt dt 0 0 ∞ ∞ d d = 1{s≤t} ϕ (t) ψ (s) f (t ∧ s)dt dt dt dt 0 0 ∞ ∞ d d ϕ (t) ψ (s) f (t ∧ s)dt dt 1{s>t} + dt dt 0 0 =: I + I I. By Fubini’s theorem ∞ ∞ d d I =− ϕ(s) ψ(t) ψ (s) f (s)ds, I I = − ϕ (t) f (t)dt. dt dt 0 0
15.2 Stochastics
383
Taking into account the boundary conditions f (t)ϕ(t) = 0, t ∈ {0, ∞}, we integrate by parts and obtain ∞ ∞ d d II = ϕ(t) ϕ(t) f (t) ψ(t)dt f (t)ψ(t))dt + dt dt 0 0 Adding up the representations for I and II we obtain ∞ ∞ ∞ d d d ψ (s) f (t ∧ s)dt dt = ( ϕ)(t) ϕ(t)ψ(t) f (t)dt. dt dt dt 0 0 0 We may rewrite (15.112) in terms of the δfunction with support in t = 0 as follows: ∞ ⎫ d d d ⎪ Covω ϕ, m i ψ, m j = ϕ(t)ψ(t) [m i , m j ](t)dt, ⎪ ⎬ dt dt dt 0 (15.113) ∞ ∞ ⎪ d d i ⎭ m (t), m j ds dt, i, j = 1, 2. ⎪ = ϕ(t)ψ(s)δ0 (t − s) dt ds 0 0
Consider now the random covariance for the family of square integrable j Rd valued continuous martingales martingales m ε (·, r0i ), m ε (·, r0 ), defined in (5.2) t m ε (t, r0i ) := Ŵε (r (u, r0i ), q)w(dq, du), i = 1, . . . N , 0
where the r (·, r0i ) are the solutions of (5.1) such that (4.11) holds. We assume that the associated diffusion matrix, Dε,kℓ (·, ·), is spatially homogeneous, i.e., that (5.12) holds. The mutual quadratic variations are continuously differentiable and the derivatives are bounded uniformly in all variables, i.e., there is a finite c > 0 such that d j i ess sup [m ε,k (t, r0 ), m ε,ℓ (t, r0 )] ≤ c k, ℓ = 1, . . . , d, i, j = 1, . . . , N . t,ω dt (15.114) In this case we can take S(R) instead of Cc∞ (R; R). We must show that m ε (t, r0i ) at most grows “slowly,” as t −→ ∞. To this end we employ a version of the asymptotic law of the iterated logarithm, whose proof is found in Lo`eve (1978), Sect. 41. Lemma 15.45. Let β(·) be a realvalued standard Brownian motion. Then lim sup 1 t→∞
β(t)
2t log log(t)
= 1 a.s.
(15.115) ⊔ ⊓
Corollary 15.46. There is a finite constant c¯ such that for all i = 1, . . . , N and k = 1, . . . , d m ε, (t, r0i ) lim sup 1 k ≤ c¯ a.s. (15.116) t→∞ 2t log log(t)
384
15 Appendix
Proof. By Proposition 5.2, m ε,k (·, r0i ) are onedimensional Brownian motions with variance Dε,kk (0). Hence (15.116) follows from (15.115) after a simple determinis⊔ ⊓ tic time change, provided Dε,kk (0) > 0. If Dε,kk (0) = 0, then m ε,k (·, r0i ) ≡ 0. As a consequence we obtain “slow” growth of m ε (t, r0i ) and we may choose the test functions from S(R) in the derivation of Theorem 15.44. We are ready to define the space–time correlations of our driving correlated Brownian motions. Recall that d (15.117) m ε (t, r0 ) = Ŵε (r (t, r0 ), q)w(dq, dt), dt interpreted as a generalized random process (in t) and indexed by the spatial initial condition r0 . Considering both t and r0 to be variables, the lefthand side defines a space–time random field, which is generalized in the variable t. Let ϕd , ψd ∈ S(Rd ) and ϕ, ψ ∈ S(R). If Fd+1 is a function of (r, t) such that Fd+1 ∈ S ′ (Rd+1 ), the duality between F and ϕd (r ) · ϕ(t) is defined by ∞ Fd+1 (r, t)ϕd (r )ϕ(t)dr dt. (Fd+1 , ϕd ϕ) := (15.118) −∞
The duality between arbitrary elements from S ′ (Rd+1 ) and test functions is an extension of this duality. Let us focus on two initial conditions, r0 and r˜0 , respectively. Definition 15.47. ⎫ d d ⎪ ⎪ Covd+1,ω ϕd ϕ, m ε (·, ·) ψd ψ, m ε (·, ·) ⎪ ⎪ ⎪ dt dt ⎪ kℓ ⎪ ∞ ∞ ⎪ ⎬ = ϕd (r0 )ψd (˜r0 )ϕ(t)ψ(s)δ0 (t − s) ⎪ 0 0 ⎪
⎪ d ⎪ ⎪ ⎪ × Ŵε,km (r (t, r0 ), q)Ŵε,ℓm (r (s, r˜0 ), q)dq ds dt dr0 d˜r0 ⎪ ⎪ ⎭
(15.119)
The definition of the δfunction implies ⎫ d d ⎪ ⎪ Covd+1,ω ϕd ϕ, m ε,k (·, ·) ψd ψ, m ε,ℓ (·, ·) ⎪ ⎪ ⎪ dt dt ⎪ kℓ ⎪ ∞ ⎪ ⎬ = ϕd (r0 )ψd (˜r0 )ϕ(t)ψ(t) ⎪ 0 ⎪
⎪ d ⎪ ⎪ ⎪ × Ŵε,km (r (t, r0 ), q)Ŵε,ℓm (r (t, r˜0 ), q) dq dt dr0 d r˜0 . ⎪ ⎪ ⎭
(15.120)
m=1
will be called the “random covariance” of the random field
d dt m ε (·, ·).
⊔ ⊓
m=1
It follows that the space–time random field of correlated Brownian motions is generalized in the time variable and “classical” in the space variable. Taking the
15.2 Stochastics
385
mathematical expectation in (15.120) (or (15.119)) yields the usual covariance as a bilinear functional on the test functions: ⎫ d d ⎪ ⎪ ECovd+1,ω ϕd ϕ, m ε,k (·, ·) ψd ψ, m ε,ℓ (·, ·) ⎪ ⎪ dt dt ⎪ kℓ ⎪ ⎪ ⎪ ⎪ ∞ ⎬ (15.121) = ϕd (r0 )ψd (˜r0 )ϕ(t)ψ(t) ⎪ ⎪ 0" ⎪ # ⎪ d ⎪
⎪ ⎪ × E Ŵε,km (r (t, r0 ), q)Ŵε,ℓm (r (t, r˜0 ), q) dq dt dr0 d r˜0 . ⎪ ⎪ ⎭ m=1
Recall the proof of Theorem 14.2 and suppose that the diffusion matrices are spatially homogeneous and independent of t and µ. Let β(·, r˜0 ) be a family of Rd valued Brownian motions with starts in r0 and covariance matrices D0 Id Brownian motions such that β(·, r0 ) and β(·, r˜0 ) are independent whenever r0 != r˜0 . It follows from (15.113) that its covariance is59 D0 δ0 (t − s) ⊗ δ0,d (r0 − r˜0 )δkℓ . By the proof of Theorem 14.2 (cf. (14.48)) we have (rε (·, r0 ), rε (·, r˜0 )) ⇒ (β(·, r0 ), β(·, r˜0 )), as ε → 0 . Thus, we obtain d d ··, m ε,ℓ (·, ·) ··, m ε,k (·, ·) dt dt kℓ ≈ D0 δ0 (t − s) ⊗ δ0,d (r0 − r˜0 )(··)δkℓ for small ε.
ECovd+1,ω
(15.122)
“··” in (15.122) is the space for the variables from S(Rd ) and S(R), respectively. However, we do not claim that β(·, ·) is a generalized ddimensional random field with covariance from the righthand side of (15.122), because we did not specify the distribution in the spatial variable. Instead of constructing β(·, ·) as a generalized random field recall the definition of the Rd valued space–time standard Gaussian white noise w(dq, dt) = (w1 (dq, dt), . . . , wd (dq, dt))T . Let us focus on the first component of w(dq, dt). Following Walsh (1986), we defined in Definition 2.2 w1 (dq, dt, ω) as a finitely additive signed measure on the Borel sets A of B d+1 of finite Lebesgue measure and as a family of Gaussian random variables, indexed by (d +1)dimensional Borel sets. We also pointed out that the definition of w1 (dq, dt) is a multiparameter generalization of the properties of a scalarvalued (standard) Brownian motion β(·) as a finitely additive signed random measure, β(dt, ω), on onedimensional Borel sets and as a family of Gaussian random variables indexed by onedimensional Borel sets. The analogy can be extended further. Recall that the generalized derivative of β(·) is a Schwarz distribution, i.e., 59
Cf. also (5.6).
386
15 Appendix
d β(·, ω) ∈ S ′ (R). dt t If we now fix t, then 0 w(dq, ds, ω) defines a finitely additive signed measure on the Borel sets B of B d of finite Lebesgue measure. By (15.69), in addition to (15.32) t and (15.36), we can interpret the finitely additive signed measure on 0 w1 (d·, ds, ω) as an element from S ′ (Rd ). Set t (15.123) 1 Br w1 (dq, ds, ω), wˆ 1 (r, t, ω) := 0
where Br is a closed rectangular domain in Rd containing 0 and with side lengths r1 , . . . , rd and whose sides are parallel to the axes. If all endpoints of Br have nonnegative coordinates, wˆ 1 (·, ·) has a continuous modification, which is called the Brownian sheet (cf. Walsh (loc. cit.)). We will assume that the left hand side of (15.123) is already the Brownian sheet. Hence, we may interpret the integration d+1 as a multiparameter integration against the Brownian sheet of a Borel set in R+ wˆ 1 (r, t). We extend the notion of a Brownian sheet to wˆ 1 (r, t, ω) for r ∈ Rd by 2d i.i.d. Brownian sheets and patching the finitely additive measure taking ∞ 1 A (q, t)w1 (dq, dt, ω) together as the sum of 2d measures, determined by 0 the multiparameter integration against 2d i.i.d. Brownian sheets, defined on each of the 2d domains Dk , k = 1, . . . , 2d , where any of the coordinates is either nonnegative or negative. Denote these Brownian sheets by wˆ 1,k , k = 1, . . . , 2d , and observe that the sum of independent normal random variables is normal and that the mean and variance of the sum is the sum of the means and variances. Therefore, we obtain
0
∞
2
d
1 A (q, t)w1 (dq, dt) ∼
k=1 0
∞
1 A∩Dk 1 A (q, t)wˆ 1,k (dq, dt), (15.124)
i.e., both sides in (15.124) are equivalent in distribution. This construction is / Dk , yields a multiparameter relatively simple and, setting wˆ 1,k (r, t) ≡ 0 for r ∈ continuous random field60 d
¯ˆ 1 (r, t) := w
2
k=1
wˆ 1,k (r, t)1 Dk (r ).
(15.125)
¯ˆ 1 (r, t, ω) in the generalized sense with respect to Hence, we can differentiate w all coordinates. Choosing ϕ(·) ∈ S(R) and ϕd (·) ∈ S(Rd ) we obtain61 60
61
Notice that the extension of the Brownian sheet to negative coordinates is analogous to the construction of a Brownian motion on R, which can be done by taking two i.i.d. Brownian motions β+ (·) and β− (·) for the Borel sets in [0, ∞) and (−∞, 0], respectively. The t time derivative leads to a Schwarz distribution by the previous arguments, because 0 ϕd (q)w(dq, ds) is a onedimensional Brownian motion.
15.2 Stochastics
0
∞
ϕd (q)ϕ(s)
387
∂ d+1 ¯ˆ 1 (q, s, ω)dq ds = w ∂s∂q1 . . . ∂qd
0
∞
ϕ(s)ϕd (q)w1 (dq, ds, ω).
We conclude62 ∂ d+1 ¯ˆ 1 (·, ·, ω) ≡ w1 (d·, d·, ω) in S ′ (Rd+1 ). w ∂s∂q1 . . . ∂qd
(15.126)
The generalization to w(dq, dt) = (w1 (dq, dt), . . . wd (dq, dt))T is carried out componentwise. Altogether, we obtain that ∂ d+1 ¯ˆ ·) ≡ w(d·, d·) w(·, ∂s∂q1 . . . ∂qd is a generalized Rd valued random field whose components take values S ′ (Rd+1 ) ¯ˆ ·) is an appropriately defined Rd valued Brownian sheet with parameter where w(·, domain Rd × [0, ∞). Its covariance is given by ⎫ ¯ˆ k (·, ·))(ψd ψ, w ¯ˆ ℓ (·, ·))]kℓ ⎪ ECovd+1,ω [(ϕd ϕ, w ⎪ ⎪ ⎪ ⎪ ∞ ⎪ d+1 ⎪ ∂ ⎪ ⎪ ¯ˆ k (q, s)dq ds ⎪ = ECovd+1,ω ϕd (q)ϕ(s) w ⎪ ⎪ ∂s∂q . . . ∂q 1 d ⎪ 0 ⎪ ⎪ ⎪ ∞ ⎪ d+1 ⎬ ∂ ¯ˆ ℓ (r, t)dr dt × ψd (r )ψ(t) w ∂t∂r1 . . . ∂rd 0 ⎪ ⎪ ⎪ ∞ ∞ ⎪ ⎪ ⎪ = ϕd (q)ϕ(s)ψd (r )ψ(t)δ0 (t − s) ⊗ δ0,d (q − r )δkℓ dq ds dr dt ⎪ ⎪ ⎪ ⎪ 0 0 ⎪ ⎪ ⎪ ∞ ⎪ ⎪ ⎪ ⎪ = ϕd (q)ψd (q)ϕ(t)ψ(t)dt dq δkℓ . ⎭ 0
(15.127) We conclude that the covariance operator of Rd valued standard Gaussian space– time white noise on Rd × R+ is δ0 (t − s) ⊗ δ0,d (q − r )δkℓ . Apart from the factor D0 this is the same covariance as for the “random field” β(·, ·). That this coincidence is not accidental follows from Remark 3.10 at the end of Chap. 3.
15.2.5 Stochastic Itˆo Integrals We summarize the construction of the Itˆo integral63 in Rd driven by finitely and infinitely many i.i.d Brownian motions. The Itˆo and the ItˆoWentzell formulas are 62
63
In the identification of a measure with a formal Radon–Nikodym derivative of that measure with respect to the Lebesgue measure as as generalized function or generalized field the “Radon– Nikodym derivative” is usually written as a suitable (partial) derivative in the distributional sense. We remind the reader of the identification the Dirac delta function with support in some point a ∈ Rd and the point measure δa (dr ). Cf. also Schauml¨offel (1986). Cf. Itˆo (1944).
388
15 Appendix
presented. Finally, we comment on the functional analytic interpretation of the Itˆo integral and provide a sufficient condition for uniform tightness of SODEs. Let β1 (·) a onedimensional standard Brownian motion and φ(·) a realvalued dt ⊗ dP square integrable adapted process on [0, T ] × Ω. We have seen in (15.115) of the previous section that, under the assumption φ(·) is a.s. continuous and of bounded variation, we may define the stochastic integral as a Stieltjes integral plus a boundary term: t t (15.128) φ(s)β1 (ds) := − β1 (s)φ(ds) + φ(t)β(t) ∈ [0, T ] 0
0
The problem of stochastic integration, however, is that most interesting integrands will not be of bounded variation. An easy way to understand this claim is considering a stochastic ordinary differential equation (SODE): dx = b(x)β1 (dt), x(0) = x0 ,
(15.129)
where we may, for the time being, assume that the initial condition is deterministic. Following the procedure of ordinary differential equations (ODEs),64 we convert (15.129) into an equivalent stochastic integral equation: t (15.130) b(x(s))β1 (ds). x(t) = x0 + 0
Assuming that the coefficient b(x) = bx with b != 0 ∀x, we again resort to the methods of ODE, apply the PicardLindel¨of procedure (as b(x) = bx is obviously Lipschitz) and try to solve (15.130) through iteration, defining recursively an approximating sequence as follows: t (15.131) x1 (t) = x0 + bx0 β1 (ds) = x0 + bx0 β1 (t), 0
The definition of the first step x1 (·) is trivial. However, we see that x1 (·) − x0 ≡ bβ1 (t), which is itself of unbounded variation and with quadratic variation [β1 ](t) ≡ t by (15.91). Therefore, we cannot explain the righthand side of the second step, presented in the following (15.132), pathwise as a Stieltjes integral through integration by parts as in (15.108): t t bx1 (s)β1 (ds) = b{bβ1 (s) + x0 }β1 (ds). (15.132) x2 (t) = x0 + 0
0
Being unable to define the stochastic integral on the righthand side of (15.132) through integration by parts need not prevent us from trying the Riemann–Stieltjes idea directly. To this end, as in Definition 15.33, consider the sequence of partitions {t0n < t1n < · · · < tkn < · · ·} and let γ ∈ [0, 1]. Set 64
Cf., e.g., Coddington and Levinson (loc.cit.).
15.2 Stochastics
389
3
n n n Sn (t, γ ) := b β1 (tk−1 ∧ t) + x0 + γ (β1 (tk ∧ t) − β1 (tk−1 ∧ t)) k≥0 (15.133) n ∧ t)), ×(β1 (tkn ∧ t) − β1 (tk−1
i.e., we choose the integrand at a fixed linear combination of the endpoints of the interval. By linearity of the summation, ⎫ Sn (t, γ ) ⎪ ⎪ ⎪ ⎪
⎪ ⎪ n n 2 ⎪ = bγ (β1 (tk ∧ t) − β1 (tk−1 ∧ t)) ⎪ ⎪ ⎪ ⎬ k≥0 (15.134)
n n ⎪ b{β1 (tk−1 ∧ t) + x0 }(β1 (tkn ∧ t) − β1 (tk−1 ∧ t)) ⎪ + ⎪ ⎪ ⎪ ⎪ k≥0 ⎪ ⎪ ⎪ ⎪ ⎭ =: Sn,1,γ (t) + Sn,2,γ (t). By (15.90) for every t ≥ 0
Sn,1,γ (t) −→ bγ [β1 ](t) ≡ bγ t a.s., as n −→ ∞ . By the proof of Proposition 15.34, t Sn,2,γ (t) −→ b (β1 (s) + x0 )β1 (ds) in probability, as n −→ ∞ .
(15.135)
(15.136)
0
Altogether, Sn (t, γ ) −→ b
t 0
(β1 (s) + x0 )β1 (ds) + bγ t in probability, as n −→ ∞, (15.137)
i.e., the limit depends on the choice of the evaluation of the integrand in the approximating sum. We conclude that we cannot use the Stieltjes integral in the PicardLindel¨of approximation of (15.130). If we choose the left endpoint in the above approximation, i.e., γ = 0, we obtain the stochastic Itˆo integral. In what follows, we provide more details about Itˆo integration. To avoid a complicated notation, we will only sketch the construction of the Itˆo integral, driven by Brownian motions, and leave it to the reader to consult the literature about the more general case of martingaledriven stochastic integrals. Set ⎫ ⎧ T d ⎬ ⎨
2 E φkℓ (t, ·)&∞ ∀T '0 , L 2,F ,loc ([0, ∞) ×  : Md×d ) := φ : ⎭ ⎩ 0 k,ℓ=1
(15.138)
390
15 Appendix
where φ(·, ·) is an Ft adapted Md×d valued process, jointly measurable in (t, ω) with respect to dt ⊗ P(dω). Further, let β(·) be an Rd valued Ft adapted standard Brownian motion. Generalizing the idea of Riemann and Stieltjes, we first define the stochastic Itˆo integral for step function type processes such that the resulting integral becomes a square integrable martingale. Suppose for φ ∈ L 2,F ,loc ([0, ∞) ×  : ˜ n) Md×d ) there is a sequence of points {0 = t0n < t1n < · · · < tkn < · · ·} and φ(t k Ftkn adapted Md×d valued random variables, i = 0, 1, . . ., such that ˜ 0n ) + φn (t, ω) = φ(t
∞
k=1
n ˜ k−1 n ,t n ] (t) ∀t a.s. φ(t )1(tk−1 k
(15.139)
φn (·, ·) is called “simple.” The stochastic (Itˆo) integral of a simple process with respect to β(·) is defined by
t
0
˜ φn (s)β(ds) := φ(0) +
∞
k=1
n n ˜ k−1 φ(t )(β(tkn ∨ t) − β(tk−1 ∨ t)) ∀t a.s. (15.140)
· As in Sect. 15.6.2, we see that 0 φn (s)β(ds) is a continuous square integrable martingale. Further, it is proved that the class of simple processes is dense in L 2,F ,loc ([0, ∞) ×  : Md×d ) (cf., e.g., Ikeda and Watanabe (loc.cit.), Chap.II.1 or Liptser and Shiryayev (1974), Chap. 4.2 as well as Itˆo (1944)). A simple application of Doob’s inequality (Theorem 15.32), as for Sn,2 (·) in theproof of Propo· sition 15.34, entails that we may construct a Cauchy sequence 0 φn (s)β(ds) in L 0,F (C([0, T ]; Rd )), as φn (·) approaches φ(·) in L 2,F ,loc ([0, ∞) ×  : Md×d ). The resulting limit is itself a square continuous Rd valued martingale and is defined as the limit in probability of the approximating stochastic (Itˆo) integrals.65 Definition 15.48. The stochastic Itˆo integral of φ(·) driven by β(·) is defined as the stochastic limit in L 0,F (C([0, T ]; Rd )) · t φn (s)β(ds), (15.141) φ(s)β(ds) = lim n→∞ 0
0
where the φn are simple processes. We verify that
d t
ℓ=1
0
⊔ ⊓
φk,ℓ (s)βℓ (ds) are continuous square integrable realvalued
martingales for k = 1, . . . , d, where φk,ℓ are the entries of the d × d matrix Φ and βℓ (·) are the onedimensional components of β(·). Further, the mutual quadratic variation of these martingales satisfies the following relation: t 3 t 2 t φk,ℓ (s)ds, if ℓ = ℓ˜ , 0 φk,ℓ (s)βℓ (ds), φk,ℓ˜(s)βℓ˜(ds) := ˜ 0, if ℓ != ℓ. 0 0 (15.142) 65
Since β(·) is continuous, we do not need to take the left hand limit in the integrand.
15.2 Stochastics
391
Expression (15.142) can be shown through approximation with stochastic Itˆo integrals of simple processes, employing (15.95), since the onedimensional Brownian motions βℓ (·) are independent and, therefore, also uncorrelated. The latter fact implies that the approximating Itˆo integrals are uncorrelated. Consequently, the t quadratic variation of 0 φ(s)β(ds) is given by
0
·
d t
φ(s)β(ds) (t) = φk,ℓ (s)ds,
(15.143)
k,ℓ=1 0
We generalize the above construction to an infinite sequence of stochastic Itˆo integrals. Let φn (·, ·) be a sequence of Ft adapted jointly measurable Md×d valued processes and set L 2,F ([0, T ] ×  × N : Md×d ) := {(φ1 , . . . φn . . .) :
d ∞
n=1 k,ℓ=1 0
T
2 Eφn,kℓ (t, ·) < ∞}.
(15.144) Further, as in (4.14), let β n (·) a sequence of Rd valued Ft adapted i.i.d. standard Brownian motions. As in (15.142), we see that t t n m (15.145) φn (s)β (ds), φm (s)β (ds) = 0, if n != m 0
0
Hence, m(t) :=
n
t
φn (s)β n (ds)
0
(15.146)
is a continuous square integrable Rd valued martingale with quadratic variation % $ d t
· n 2 (15.147) φn (s)β (ds) (t) ≡ φn,kℓ (s)ds 0
n
n kℓ=1 0
and, by Doob’s inequality, 2 d T t
n 2 E sup φn (s)β (ds) ≤ 4 Eφn,kℓ (s)ds ∀T > 0. 0 0 0≤t≤T n
n kℓ=1
(15.148)
Remark 15.49.
(i) The construction of Itˆo integrals with respect to a series of uncorrelated square integrable martingales follows exactly the same pattern. Similarly, the theory immediately generalizes to semimartingales of the following form
392
15 Appendix
a(t) := b(t) + m(t),
(15.149)
where b(·) is a process of bounded variation and m(·) is a square integrable martingale. (ii) The following generalization of (15.147) can be derived and is often quite useful. Suppose φ(·) ∈ L 2,F ,loc ([0, ∞) × ; Md×d ) satisfies d
T
i, j=1 0
Eφi2j (s)[m i ](ds) < ∞ ∀T,
(15.150)
where m(·) = (m 1 (·), . . . , m d (·))T is a continuous square integrable Rd valued martingale. The Itˆo integral t φ(s)m(ds) 0
is defined similarly to the Itˆo integral, driven by Brownian motions. It is also a continuous square integrable Rd valued martingale and
0
·
t d
φ(s)m(ds) (t) ≡ φi j (s)φik (s)[m j , m k ](ds).
(15.151)
i, j,k=1 0
(iii) Using stopping times (15.150) can be relaxed to d
i, j=1 0
T
φi2j (s)[m i ](ds) < ∞
a.s. ∀T,
(15.152)
t We obtain that 0 φ(s)m(ds) exists and is a continuous locally square integrable Rd valued martingale. The extension of this statement to stochastic integrals, driven by continuous semimartingales, is obvious. (iv) Apart from the necessary integrability assumptions on the integrand the most n ) important assumption in stochastic Itˆo integration is that the integrand φ(tk−1 n n n adapted and that the integrator increments m(t ) − m(t must be Ftk−1 k k−1 ) are n orthogonal (i.e. uncorrelated) to φ(tk−1 ). We abbreviate this property by66 n n ) φ(tk−1 ) ⊥ m(tkn ) − m(tk−1
∀n, k
and, for the limits,φ(s) ⊥ m(ds) ∀s. (15.153) ⊔ ⊓
The Itˆo integration has been extended to integrals driven by local (cadlag) martingales.67 In fact, the Itˆo integral may be defined for general “good integrators,” which are stochastic processes and allow a version of Lebesgue’s dominated convergence 66 67
For discontinuous martingales we need to evaluate the integrand at s− in (15.153). Cf., e.g., Metivier and Pellaumail (1980), Liptser and Shiryayev (1974, 1986) as well as Protter (2004).
15.2 Stochastics
393
theorem.68 One then shows that these good integrators are semimartingales, i.e., can be represented as in (15.149), where m(·) only needs to be integrable. Such a representation is called the “DoobMeyer decomposition” of a semimartingale. We will return to the “good integrator” approach at the end of this subsection. Next, we present the Itˆo formula, which is the most important formula in stochastic analysis and is the key in the transition from a Brownian particle movement to secondorder parabolic PDEs and, as shown in Chaps. 8 and 14, also to secondorder parabolic SPDEs and their macroscopic limits.69 Again, we want to avoid a cumbersome notation, and will only present a special case, which was used in the derivation of SPDEs in this volume. Theorem 15.50. Itˆo’s Formula: First Extended Chain Rule Let ϕ(r, t) be a function from Rd+1 into R. Suppose ϕ is twice continuously differentiable function with respect to the spatial variables and once continuously differentiable with respect to t such that all partial derivatives are bounded. Let m(·) be a continuous square integrable Rd valued martingale and b(·) a continuous process of bounded variation. We set a(t) := b(t) + m(t), (cf. (15.149)). ϕ(a(·), t) is a continuous, locally square integrable semimartingale and the following formula holds: ⎫ t ∂ ⎪ ⎪ ϕ(a(t), t) = ϕ(a(0), 0) + ϕ (a(s), s)ds ⎪ ⎪ ⎪ ∂s 0 ⎪ ⎪ ⎪ t ⎪ ⎪ ⎬ + (∇ϕ)(a(s), s) · (b(ds) + m(ds)) (15.154) 0 ⎪ ⎪ ⎪ ⎪ d ⎪ ⎪ 1 t 2 ⎪ ⎪ + (∂i, j ϕ)(a(s), s)[m i , m j ](ds), ⎪ ⎪ ⎭ 2 0 i, j=1
where [m i , m j ](·) are the mutual quadratic variations of the onedimensional components of m(·).
Proof. (Sketch) We only sketch the proof and refer the reader for more details to any book on stochastic analysis.70 Let {t0n < t1n < · · · < tkn < · · ·} be a sequence of partitions of [0, ∞) as in Definition 15.33. Then, 68 69 70
Cf. Protter (loc.cit.), Chap. IV, Theorem 32. Recall from Chap. 14 that these macroscopic limits are themselves solutions of secondorder parabolic PDEs. Cf., e.g., Ikeda and Watanabe (loc.cit.), Chap. II.5 or Gikhman and Skorokhod (1982), Chap.3.2.
394
15 Appendix
ϕ(a(t), t) − ϕ(a(0), 0)
n n = ϕ(a(tkn ∧ t), tkn ∧ t) − ϕ(a(tk−1 ∧ t), tk−1 ∧ t) k
=
n n n ϕ(a(tk−1 ∧ t), tkn ∧ t) − ϕ(a(tk−1 ∧ t), tk−1 ∧ t)
⎫ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎬
⎪ ⎪ ⎪ ⎪ ⎪ ⎪
⎪ n n n n ⎪ + ϕ(a(tk ∧ t), tk ∧ t) − ϕ(a(tk−1 ∧ t), tk ∧ t) ⎪ ⎪ ⎪ ⎪ ⎪ k ⎪ ⎪ ⎪ ⎭ =: Sn,1 (t) + Sn,2 (t). k
(15.155)
Apparently, t ∂ ϕ(a(s), s)ds, as n −→ ∞, uniformly on compact intervals [0, T ]. Sn,1 (t) −→ ∂t 0 (15.156) By Taylor’s formula,
⎫ n ⎪ Sn,2 (t) = ∧ t), tkn ∧ t) ϕ(a(tkn ∧ t), tkn ∧ t) − ϕ(a(tk−1 ⎪ ⎪ ⎪ ⎪ k ⎪ ⎪ ⎪
⎪ n n n n ⎪ = ∇ϕ(a(tk−1 ∧ t), tk ∧ t) · (a(tk ∧ t) − a(tk−1 ∧ t)) ⎪ ⎪ ⎪ ⎪ ⎪ k ⎪ ⎬ d (15.157)
1
⎪ n n ⎪ + ∂i,2 j ϕ(a(tk−1 ∧ t), tk1 ∧ t)(ai (tkn ∧ t) ⎪ ⎪ ⎪ 2 ⎪ k i, j=1 ⎪ ⎪ ⎪ ⎪ ⎪ n n n ⎪ −ai (tk−1 ∧ t))(a j (tk ∧ t) − a j (tk−1 ∧ t)) ⎪ ⎪ ⎪ ⎪ ⎭ +Rn (t, ϕ, a),
where the remainder term Rn (t, ϕ, a) tends to 0 in probability.71 Next, we note that n n ∧ t) ∀n, k. ∧ t), tkn ) ⊥ m(tkn ∧ t) − m(tk−1 ∇ϕ(a(tk−1
(15.158)
n ∧t), t n ) satisfies the assumptions of the construction of the Therefore, ∇ϕ(a(tk−1 k n ∧ t). Itˆo integral with respect to the semimartingale increments a(tkn ∧ t) − a(tk−1 n n Hence, the integrability assumptions on ∇ϕ(a(tk−1 ∧ t), tk ) allow us to conclude
71
Assuming more differentiability on ϕ, we may show that the remainder term is dominated by n n n (ai (tkn ∧ t) − ai (tk−1 ∧ t))(a j (tkn ∧ t) − a j (tk−1 ∧ t))(aℓ (tkn ∧ t) − aℓ (tk−1 ∧ t)). As our n
martingale has finite quadratic variation, this sum has to tend to zero. This is very similar to showing that the quadratic variation of a continuous process of bounded variation equals 0.
15.2 Stochastics
k
395
⎫ n n ∇ϕ(a(tk−1 ∧ t), tkn ∧ t) · (a(tkn ∧ t) − a(tk−1 ∧ t)) ⎪ ⎪ ⎪ ⎪ ⎬ t
∇ϕ(a(s), s) · (b(ds) + m(ds)) as n −→ ∞, ⎪ ⎪ ⎪ 0 ⎪ ⎭ uniformly on compact intervals [0, T ],
−→
(15.159)
Finally, recall that, by (15.93) we have for the mutual quadratic variations [ai , a j ](t) = [m i , m j ](t).
(15.160)
Therefore, the integrability assumptions imply d
1
k
2
i, j=1
n ∂i,2 j ϕ(a(tk−1
n ∧ t), tk−1
∧ t)(ai (tkn
⎫ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎬
∧ t)
n n −ai (tk−1 ∧ t)) ∧ t))(a j (tkn ∧ t) − a j (tk−1 ⎪ d ⎪ 1 t 2 ⎪ ⎪ −→ ∂i, j ϕ(a(s), s)[m i , m j ](ds), as n −→ ∞, ⎪ ⎪ ⎪ 2 ⎪ ⎪ i, j=1 0 ⎪ ⎪ ⎪ ⎭ uniformly on compact intervals [0, T ],
where the righthand side is the usual (Lebesgue) Stieltjes integral.
(15.161)
⊔ ⊓
As an application of the Itˆo formula we now give the proof of Theorem 15.37 (Levy’s characterization of Brownian motion), following Ethier and Kurtz (loc.cit.), Chap. 7.1, Theorem 1.1. Proof of Theorem 15.37 Let θ ∈ Rd be arbitrary and set 1 ϕ(r, t) := exp[iθ · r + θ · (Cθ )t]. 2 where i :=
3
√ −1. Applying the Itˆo formula we obtain 1 exp iθ · m(t) + θ · (Cθ )t 2 t 1 1 θ · (Cθ )ds exp iθ · m(s) + θ · (Cθ )s =1+ 2 2 0 t 1 + exp iθ · m(s) + θ · (Cθ )s iθ · m(ds) 2 0 t 1 1 − exp iθ · m(s) + θ · (Cθ )s θ · (Cθ )ds. 2 0 2
(15.162)
396
15 Appendix
Hence, ⎫ 1 ⎪ ⎪ exp iθ · m(t) + θ · (Cθ )t ⎪ ⎬ 2 t ⎪ 1 ⎪ ⎭ exp iθ · m(s) + θ · (Cθ )s iθ · m(ds). ⎪ =1+ 2 0
(15.163)
The righthand side is (an Ft −) continuous square integrable complexvalued martingale. Therefore, for 0 ≤ s < t 3 1 1 E exp iθ · m(t) + θ · (Cθ )t Fs = exp iθ · m(s) + θ · (Cθ )s 2 2
or, equivalently, 3 1 1 E exp iθ · (m(t) − m(s)) + θ · (Cθ )t Fs = exp θ · (Cθ )(t − s) . 2 2 (15.164) Consequently, m(·) is an Rd valued Brownian motion with covariance Ct. ⊔ ⊓
There are many extensions of Itˆo’s formula in finite dimensions as well as extensions to Hilbert space valued semimartingales.72 In fact, (15.89) is a special case of the Itˆo’s formula in Hilbert space. For finite dimensional more general semimartingales we refer the reader to Ikeda and Watanabe (loc.cit.) and Gikhman and Skorokhod (loc.cit.) Chap. 4.4. Krylov (1977), Chap. II.10, proves an Itˆo’s formula in finite dimensions where the function ϕ only needs to have second derivatives in the generalized sense in addition to some other conditions. The semimartingales in Krylov’s Itˆo formula are continuous and represented as solutions of Itˆos SODEs driven by Brownian motions. We now present a generalization in finite dimensions, which applies to continuous semimartingales but where the function ϕ is replaced by a semimartingale that depends on a spatial parameter.73 Theorem 15.51 ItˆoWentzell Formula – Second Extended Chain Rule Consider the semi–martingale from (7.5) t t ˆS(r, t) := F(r, u)du + J (r, p, u)w(d p, du). 0
(15.165)
0
Suppose F is continuously differentiable in r and J (r, p, u) is twice continuously differentiable in r . Let a(·) be a continuous locally square integrable Rd valued semimartingale with representation (15.152), a(·) = b(·) + m(·). Deˆ then the following formula holds for ℓ = noting by Sˆℓ the ℓth coordinate of S, 1, . . . , d: 72 73
Cf. for the latter, Metivier and Pellaumail, loc.cit. Chap. 2. Cf. Kunita, loc.cit., Chap. 3.3, Thereom 3.3.1, Generalized Itˆo formula. In view of the comparison of our approach with Kunita’s formalism from our Chap. 7, we use our notation.
15.2 Stochastics
397
t
Sˆℓ (a(t), t) = Sˆℓ (a(0), 0) + Sˆℓ (a(s), ds) 0 t + (∇ Sˆℓ )(a(s), s) · (b(ds) + m(ds)) 0
⎫ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎬
d 1 t 2 ˆ ⎪ + (∂i, j Sℓ )(a(s), s)[m i , m j ](ds) ⎪ ⎪ ⎪ 2 0 ⎪ i, j=1 ⎪ ⎪ ⎪ t
d ⎪ ⎪ ∂ ⎪ ⎪ Jℓ, j (a(s), p, s)w j (d p, ds), m i (ds) . ⎪ + ⎪ ⎭ ∂r i 0
(15.166)
i, j=1
Proof. (Sketch) We again only sketch the proof and refer the reader for more details to Kunita (loc.cit.) for a short proof and to Rozovsky (1983), Chap. 1.4, Theorem 9, for a ˆ t) has a somewhat different represendetailed proof, where the semimartingale S(r, tation. We basically repeat the proof of the Itˆo formula with Sˆℓ instead of ϕ. second term on the righthand side of (15.166) is formally equivalent to The t ∂ ˆ 0 ( ∂s Sℓ )(a(s), s)ds, which corresponds to the second term on the righthand side of (15.154). We need to be more careful when applying Taylor’s formula to Sn,2 (t). n The problem is that, unlike in (15.158),74 ∇ Sˆℓ (a(tk−1 ∧ t), tkn ∧ t) is anticipating n n with respect to the increments (a(tk ∧t)−a(tk−1 ∧t)) and we cannot expect the sum over all k to converge to an Itˆo integral driven by the semimartingale increments a(ds). Therefore, we must “correct” this term first before we employ martingale sums and Doob’s inequality as in Proposition 15.34. We do this as follows: ⎫ n n ∇ Sˆℓ (a(tk−1 ∧ t), tkn ∧ t) · (a(tkn ∧ t) − a(tk−1 ∧ t)) ⎪ ⎪ ⎬ n n n n ∧ t)) · (a(tkn ∧ t) − a(tk−1 ∧ t)) = (∇ Sˆℓ (a(tk−1 ∧ t), tkn ∧ t) − ∇ Sˆℓ (a(tk−1 ∧ t), tk−1
n n n +∇ Sˆℓ (a(tk−1 ∧ t), tk−1 ∧ t) · (a(tkn ∧ t) − a(tk−1 ∧ t)).
⎪ ⎪ ⎭
(15.167) We now have the analogue of (15.161), namely, n n n ∇ Sˆℓ (a(tk−1 ∧ t), tk−1 ∧ t) ⊥ m(tkn ∧ t) − m(tk−1 ∧ t),
(15.168)
whence summing up over all k, the integrability assumptions imply ⎫
n n n ∧ t)) ⎪ ∧ t) · (a(tkn ∧ t) − a(tk−1 ∧ t), tk−1 ∇ Sˆℓ (a(tk−1 ⎪ ⎪ ⎪ ⎪ k ⎪ ⎬ t
−→
0
∇ Sˆℓ (a(s), s) · (b(ds) + m(ds))
as n −→ ∞, uniformly on compact intervals [0, T ]. 74
ϕ(a(s), t) is Fs adapted for all t, because ϕ(r, t) is deterministic.
⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎭
(15.169)
398
15 Appendix
ˆ ·) are themselves Next, our assumptions imply that the partial derivatives of S(r, continuous semimartingales. Therefore,
n n n n ∇ Sˆℓ (a(tk−1 ∧ t), tkn ∧ t)−∇ Sˆℓ (a(tk−1 ∧ t), tk−1 ∧ t) · (a(tkn ∧ t) − a(tk−1 ∧ t) k
converges toward the mutual quadratic variation of the partial derivatives of ˆ ·) and the semimartingale a(·). Further, observe that, by (15.193), the terms S(r, of continuous bounded variation do not contribute to the limit. Thus, we obtain the “correction term” to the classical Itˆo formula ⎫ t
d ⎪ ∂ ⎪ Jℓ, j (a(s), p, s)w j (d p, ds), m i (ds) ⎪ ⎪ ⎪ ⎬ ∂ri 0 i, j=1 (15.170) ⎪ ∞ t
d
∂ ⎪ n ⎪ = σn,ℓ, j (a(s), s)β j (ds), m i (ds) , ⎪ ⎪ ⎭ ∂ri 0 n=1
i, j=1
employing for the second equality (4.15) with {β n (·)} a system of i.i.d. standard Brownian motions and the representation ∂ ∂ ⊔ ⊓ 'n ( p)d p. σn (r, t) := J (r, p, t)φ ∂ri ∂ri
Rd valued
We remark that the second representation of the correction formula in (15.170) is an extension (from a finite sum to an infinite series) of the formula proved in Rozovski (loc.cit.) We now return to the “good integrator” approach, following Protter (loc.cit.) and restricting both integrators and integrands to Md×d  and Rd valued processes, respectively. In the generality of cadlag integrators we need to have some restrictions on the integrands. The first step is to consider processes that are continuous from the left with limits from the right as integrands, called by its French acronym “caglad.” If f (·) is cadlag then the process f (·−), defined by f (t−) := lim f (s) ∀t s↑t
is caglad. Hence, we may work with cadlag processes both for integrators and integrands, taking always the lefthand limits of the integrand. The class of integrands will be denoted D([0, ∞); Md×d ).75 Our integrators will be cadlag semimartingales with decomposition (15.149). The construction of the stochastic integral · f (s−)a(ds) ∈ L 0 (; D([0, ∞); Rd ))) J (φ, a) := (15.171) 0
for f (·) ∈ L 0 (; D([0, ∞); Md×d )) and the semimartingale a(·) ∈ L 0 (; D ([0, ∞); Rd )) follows the pattern of the continuous case, which we outlined earlier.76 It is important to note that for stochastic integral driven by the martingale term, m(·), we employed Doob’s inequality (Theorem 15.32), which implies that 75 76
We endow Md×d with the usual matrix norm · Md×d . Cf. Proposition 15.34 and Definition 15.48.
15.2 Stochastics
399
the stochastic integral is the limit in probability uniformly on compact intervals. This fact is independent of whether or not m(·) is continuous or merely cadlag. For the integral driven by the process of bounded variation, b(·), we also employ the uniform metric and the total variation of the onedimensional components of b(·) to obtain similar estimates. Therefore, we need to endow both D([0, ∞); Md×d ) and D([0, ∞); Rd ) with the uniform metric du,Md×d and du,Rd , respectively. The latter is defined by ∞
du,Rd ( f, g) := sup ρ( f (t) − g(t))2−n (15.172) n=1 0≤t≤n
and similarly for du,Md×d . Further, L 0 (; D([0, ∞); Md×d )) and L 0 (; D ([0, ∞); Md×d )) are the corresponding classes of processes, endowed with the metric of convergence in probability, dprob,u,∞,Md×d and dprob,u,∞,Rd , respectively.77 Protter proves the following78 Theorem 15.52. The mapping J ((·), a) : (L 0 (; D([0, ∞); Md×d )), dprob,u,∞,Md×d ) −→ (L 0 (; D([0, ∞); Rd )), dprob,u,∞,Rd )
is continuous. We next state a convergence condition for stochastic integrals due to Kurtz and Protter (1996), Sects. 6 and 7. The semimartingale integrators take values in separable Banach space, and most properties are formulated with respect to the weak topology of the Banach space. We formulate a special case of Theorem 7.5 in Kurtz and Protter (loc.cit.), which will be sufficient for the application in the mesoscopic limit theorem. Let H be a separable Hilbert space with norm · H and scalar product &·, ·'H . Suppose m nk (·) is for each k ∈ N a sequence a realvalued square integrable mean zero martingales, n ∈ N ∪ {∞}. Let φk be a CONS for H and suppose that for each n ∈ N ∪ {∞} and ϕ ∈ H the series
m nk (·)&φk , ϕ'H k∈N
converges in mean square uniformly on compact intervals. Set
M n (·) := m nk (·)φk , n ∈ N ∪ {∞}.
(15.173)
k∈N
77 78
Cf. (15.84) with Rd or Md×d instead of H. Protter (loc.cit.), Chap. II, Theorem 11. Protter’s definition and his Theorem 11 are stated only for the case where both integrands and integrators are real valued. Since the stochastic integral is constructed for each component in the product between a matrix and a vector we may, without loss of generality, state the theorem in the appropriate multidimensional setting.
400
15 Appendix
We call M n (·) an Hvalued weak (square integrable) martingale. n Generalizing m k (·)φk , we define the considerations of Sect. 15.2.3 and 15.2.4 to the sequence k∈N
the tensor quadratic variation of M n (t) by
[M n ](t) := [m nk , m nℓ ](t)φk ⊗ φℓ . k, ℓ ∈ N.
(15.174)
φk ⊗ φℓ in (15.174) is the tensor product of φk and φℓ . We verify that the above convergence condition is equivalent to the statement that for each t E[M n ](t) is a bounded operator on H. If E[M n ](t) is nuclear for each t (i.e., it is the product of two HilbertSchmidt operators), we call M n (·) “regular.” Otherwise we call M n (·) “cylindrical.”79 Next, let J be the collection of Hvalued cadlag processes #(·) which is represented by L
#(·) = f k (·)φ˜ k , L ∈ N, φ˜ k ∈ H, (15.175) k=1
where f k (·) are realvalued cadlag processes. The stochastic integral for #(·) ∈ A and M n (·) is defined by t L t
&#(s−), M n (ds)' := f k (s−)m ℓ (ds)&φ˜ k , φℓ 'H (15.176) 0
k=1 ℓ∈N 0
The quadratic variation of 0 &#(s−), M n (ds)' can be represented by " L G # L t tF
n n f k (s−)φ˜ k &φ(s−), M (ds)' := . f k (s−)φ˜ k , [M ](ds) 0
0
k=1
k=1
H
(15.177)
Hence, by the boundedness of [M n ](t), we extend this definition to all Hvalued square integrable cadlag processes #(·), using a Fourier expansion of #(·) with respect to some CONS and obtain for the quadratic variation t t &#(s−), M n (ds)' := &[M n ](ds)(#(s−), #(s−)'H . (15.178) 0
0
Needless to say that, employing stopping times, we can extend the definition of the stochastic integral and (15.178) to the case where the tensor quadratic variation [M n ](·) is only locally integrable.
Definition 15.53. The sequence M n (·) of Hvalued (possibly cylindrical) martingales, n ∈ N is called “uniformly tight” if for each T and δ > 0 there is a constant K (T, δ) such that & B 1 t A n (15.179) P sup #(s−), M (ds) ≥ δ ≤ δ 0 0≤t≤T K (T, δ) for each n and all Fn,t adapted #(·) ∈ A satisfying sup #n (t)H ≤ 1. 0≤t≤T
79
Cf. Definition 15.28 of regular and cylindrical Hvalued Brownian motions.
⊔ ⊓
15.2 Stochastics
401 ¯
Next, we consider a sequence of stochastic integral equations with valued in Rk . The integrands are Fˆ n (ˆr , s−) = (F1n (ˆr , s−), . . . , Fk¯n (ˆr , s−))T such that Fin (ˆr , ·) ∈ ¯
H, n ∈ N ∪ {∞}. Further, instead of the usual initial condition there is an Rk valued sequence of cadlag processes U n (·), n ∈ N ∪ {∞}. Hence the sequence of stochastic integral equations can be written as t: ; n n rˆ (t) = U (t) + Fˆ n (ˆr (s−), s−), M n (ds) , n ∈ N ∪ {∞}, (15.180) 0
where
& Fˆ n (ˆr (s−), s−), M n (ds)' := (& Fˆ1n (ˆr (s−), s−), M n (ds)', . . . , & Fˆk¯n (ˆr (s−), s−), M n (ds)')T . Kurtz and Protter provide conditions for the existence of solutions and uniqueness, and we will just assume the existence of solutions of (15.180) for n ∈ N ∪ {∞} and uniqueness for the limiting case n = ∞. Further, suppose that the following condition holds on the coefficients Fˆ n : ∀c > 0, t > 0 sup sup 0 ≤ s ≤ tFin (ˆr , s) − Fi∞ (ˆr , s)H → 0, ˆr ≤c
(15.181)
¯ as n → ∞ , i = 1, . . . , k. Theorem 15.54. (i) Suppose H = H0 and M n (·), given by (15.173), is uniformly tight and that (15.181) holds. Further, suppose that for any finite {φ˜ k , k = 1, . . . , L} ⊂ H0 (U n (·), &M n (·), φ˜ 1 ', . . . , &M n (·), φ˜ L ') ⇒ (U ∞ (·), &M ∞ (·), Φ˜ 1 ', . . . , &M ∞ (·), φ˜ L ') ¯
in D([0, ∞); Rk+L ).
Then for any finite {φ˜ k , k = 1, . . . , L} ⊂ H0
(U n (·), rˆ n (·), &M n (·), φ˜ 1 ', . . . , &M n (·), φ˜ L ')
⇒ (U ∞ (·), rˆ ∞ (·), &M ∞ (·), φ˜ 1 ', . . . , &M ∞ (·), φ˜ L ')
(15.182)
¯
in D([0, ∞); R2k+L ). (ii) Fix γ > d.80 Suppose that, in addition to the conditions of part (i), for all tˆ, ǫ > 0 and L > 0 there exists a δ > 0 such that for all n P{sup &M n (·), ϕ > 'L} ≤ ǫ wheneverϕγ ≤ δ. t tˆ
(ii) Then, ¯
(U n (·), rˆ n (·), M n (·)) ⇒ (U ∞ (·), rˆ ∞ (·), M ∞ (·) in D([0, ∞); R2k × H−γ ).
(15.183)
80
Recall (15.32) and (15.36).
402
15 Appendix
Proof. Part (i) is a special case of Theorem 7.5 in Kurtz and Protter (loc.cit.) using the uniqueness of the the solution of the limiting equation. Part (ii) follows from (i) and Walsh (loc.cit.), Corollary 6.16. ⊔ ⊓ Remark 15.55. To better understand the uniform tightness condition in the paper by Kurtz and Protter, we now restrict the definitions and consequences to the finitedimensional setting of Theorem 15.51. To this end let an (·) be a sequence of Rd valued cadlag Fn,t semimartingales and f n (·) families of Md×d valued Fn,t adapted cadlag simple processes. (i) The sequence an (·) is called “uniformly tight” if for each T and δ > 0 there is a constant K (T, δ) such that for each n and all Md×d valued Fn,t adapted cadlag processes f n (·) t 1 P sup f n (s−)an (ds) ≥ δ ≤ δ provided sup f n (t)Md×d ≤ 1 . 0≤t≤T K (T, δ) 0 0≤t≤T &
(ii) It follows from Sect. 15.1.6 that (D([0, ∞); Rd ), du,Rd ) is a complete metric space and, by the vector space structure of D([0, ∞); Rd ), it is a Frech´et space.81 Hence, (L 0 (; D([0, ∞); Rd )), dprob,u,∞,Rd ) is also a Frech´et space space. We obtain the same property for (D([0, ∞); Md×d ), du,Md×d ) and (L 0 (; D([0, ∞); Md×d )), dprob,u,∞,Md×d ). Since the mapping J (·, a) for fixed a(·) is linear Theorem 15.52 asserts J (·, a) ∈ L(L 0 (; D([0, ∞); Md×d )), L 0 (; D([0, ∞); Rd ))),
(15.184)
i.e., J (·, a) is a bounded linear operator from the Frech´et space space L 0 (; D([0, ∞); Md×d )) into the Frech´et space space L 0 (; D([0, ∞); Rd )). (iii) The subindex n at f n (·) in Definition 15.53 merely signifies adaptedness to the family Fn,t . Therefore, we may drop the subindex at f n (·), and we may also use the metrics on [0, ∞) in the above definition. Further, we may obviously drop the requirement that f (·) be in the unit sphere (in the uniform metric), incorporating the norm of f (·) into the constants as long as the norm of f (·) does not change with n. The assumption that f (t, ω) be bounded in t and ω implies boundedness of (·) in the uniform metric in probability. Having accomplished all these cosmetic changes, the property that an (·) be uniformly tight implies the following: For each f (·) ∈ L 0 (; D([0, ∞); Md×d )), the set {J ( f, an ) : n ∈ N} is bounded in L 0 (; D([0, ∞); Rd )), i.e., it implies the assumptions of the uniform boundedness principle (Theorem 15.3). (iv) We expect that a similar observation also holds in the Hilbert space case if, as in the proof of Theorem 15.54, we can find, uniformly in t, a HilbertSchmidt imbedding for the integrators. ⊔ ⊓ 81
Cf. Definition 15.2. D([0, ∞); Rd ), endowed with the Skorokhod metric d D,Rd , (D([0, ∞); Rd ), d D,Rd ), is both complete and separable. Hence, (D([0, ∞); Rd ), d D,Rd ) is a separable Frech´et space.
15.2 Stochastics
403
15.2.6 Stochastic Stratonovich Integrals We briefly describe the Stratonovich integral, following IkedaWatanabe, and provide the transformation rule by which one may represent a Stratonovich integral as a sum of a stochastic Itˆo integral and a Stieltjes integral. Finally, we apply the formulas to a special case of the SODE (4.9). Recall that by (15.139) and (15.81) the Itˆo integral for continuous semimartingale integrators is constructed as in a formal Stieltjes approximation, but taking the integrand at the left endpoint of a partition interval and the integrator as the difference of both endpoints. Stratonovich (1964) takes the midpoint. ˜ + m(·) Let a(·) ˜ = b(·) ˜ and a(·) = b(·) + m(·) be continuous locally square integrable realvalued semimartingales (cf. (15.149)), adapted (as always in this book) to the filtration Ft . Further, let {t0n < t1n < · · · < tkn < · · ·} be a sequence of partitions of [0, ∞) as in Definition 15.33. Set
1 n n Sn (t, a, {a(t ˜ kn ) + a(t ˜ k−1 )}(a(tkn ) − a(tk−1 )). ˜ a) := (15.185) 2 k
Ikeda and Watanabe show that Sn (t, a, ˜ a) converges in probability, uniformly on compact intervals [0, T ]. The proof of this statement is very similar to the proof of the Itˆo formula. Thus, the following is well defined: Definition 15.56.
0
t
a(s) ˜ ◦ a(ds) := lim Sn (t, a, ˜ a) n→∞
(15.186)
is called the “Stratonovich integral of a(·) ˜ with respect to a(ds).”82
⊔ ⊓
The representation of the approximating sequence in (15.185) immediately implies the following transformation rule: Theorem 15.57. Under the above assumptions t t a(s) ˜ ◦ a(ds) ≡ a(s)a(ds) ˜ + [m, ˜ m](t) a.s., 0
(15.187)
0
where the stochastic integral in the righthand side is the Itˆo integral. Proof.
1 k
=
82
n n {a(t ˜ kn ) + a(t ˜ k−1 )}(a(tkn ) − a(tk−1 ))
2
k
n n a(t ˜ k−1 )(a(tkn ) − a(tk−1 )) +
1 k
2
⎫ ⎪ ⎪ ⎪ ⎬
n n ⎪ {a(t ˜ kn ) + a(t ˜ k−1 )}(a(tkn ) − a(tk−1 )). ⎪ ⎪ ⎭
(15.188)
n Cf. with our attempt to define a Stieltjes approximation for (15.132), choosing {β1 (tk−1 ∧ t) + n ∧ t))} as the evaluation point for the integrand β (·) + x . We saw x0 + γ (β1 (tkn ∧ t) − β1 (tk−1 1 0 that Itˆo’s choice is γ = 0, whereas, by (15.185), Stratonovich’s choice is γ = 12 .
404
15 Appendix
Clearly, the first sum tends to the stochastic Itˆo integral, whereas the second sum tends to the mutual quadratic variation of a(·) ˜ and a(·). By (15.93) this process reduces to the mutual quadratic variation of m(·) ˜ and m(·). ⊔ ⊓ The mutual quadratic variation [m, ˜ m](·) in (15.187) is usually called the “correction term”, which we must add to the Itˆo integral to obtain the corresponding Stratonovich integral. The generalization to multidimensional semimartingales follows from the real case componentwise. Although the Stratonovich integral requires more regularity on the integrands and integrators than the Itˆo integral, transformations of Stratonovich integrals follow the usual chain rule, very much in difference from the Itˆo integral where the usual chain rule becomes the Itˆo formula. We adopt the following theorem from Ikeda and Watanabe (loc.cit.), Chap. III.1, Theorem 1.3. Theorem 15.58. The Chain Rule Suppose a(·) := (a1 (·), . . . , ad (·)) is an Rd valued continuous locally square integrable semimartingale and ϕ ∈ C 3 (Rd ; R). Then, a(·) ˜ := ϕ(a(·)) is a realvalued continuous square integrable semimartingale and the following representation holds: d t
(15.189) a(t) ˜ ≡ ϕ(a(0)) + (∂i ϕ)(a(s)) ◦ ai (ds). i=1
0
⊔ ⊓
In what follows, we will apply the Stratonovich integral to the study of a special ˜ ∈ case of semimartingales, appearing in the definition of our SODEs (4.9) with Y(·) M f,loc,2,(0,T ] ∀T > 0. More precisely, we consider ⎫ ⎬ ˜ dr (t) = F(r (t), Y(t), t)dt + J (r (t), p, t)w(d p, dt) ⎭ r (0) = r0 ∈ L 2,F0 (Rd ), (Y˜ + , Y˜ − ) ∈ L loc,2,F (C((0, T ]; M × M)), (15.190)
assuming the conditions of Hypothesis 4.2, (i).83 In the notation of Kunita (cf. also our Sect. 7) (15.190) defines the following space–time field, which is a family of semimartingales, depending on a spatial parameter: t t ˜ (15.191) + 0 J (r, q, s)w(dq, ds). L(r, t) := 0 F(r, Y(s))ds
˜ Note that J does not depend on some measure Y(·). If, in addition F does ˜ not depend on the measure process Y(·), the field is Gaussian. The assumption of ˜ allows us to independence of the diffusion coefficient J of measure processes Y(t) define the Stratonovich integrals with respect to w(dq, ds) in terms of Itˆo integrals. 83
(15.190) is a special case of the SODE (4.9).
15.2 Stochastics
405
Proposition 15.59. Let z(·) be a continuous square integrable Rd valued semimartingale. Further, suppose that, in addition to Hypothesis 4.2, J (r, q, s) is twice continuously differentiable with respect to r such that supr ∈Rd
d
T
i, j,k,ℓ=1 0
(∂i2j Jk,ℓ )2 (r, q, s)dq ds < ∞.
(15.192)
Then, employing the representation (4.15), for all n the entries of J (z(·), q, ·) φˆ n (q)dq are continuous square integrable Rd valued semimartingales. Proof. We apply Itˆo’s formula (Theorem 15.50) to the functions ϕ(r, t) := Jkℓ (r, q, t)Φn (q)dq, k, ℓ = 1, . . . , d.
⊔ ⊓
Consequently, employing the series representation (4.15) and assuming the conditions of Proposition 15.59, we can define the Stratonovich integral of J (z(s), q, s) with respect to w(dq, ds) as a series of Stratonovich integrals from Definition 15.56 by84 t 0
J (z(s), q, s)w(dq, ◦ds) :=
∞ t
n=1 0
J (z(s), q, s)φˆ n (q)dq ◦ β n (ds), (15.193)
where the ith component of the stochastic integral is defined by t 3 n ˆ J (z(s), q, s)φn (q)dq ◦ β (ds) 0
:=
d t
j=1 0
i
(15.194)
Ji j (z(s), q, s)φn (q)dq ◦ β nj (ds).
Next, we consider the solution of (15.190). Since the coefficients are bounded this solution is a continuous square integrable Rvalued semimartingale. Proposition 15.60. Let r (·) be the solution of (15.190). Suppose that, in addition to the conditions of Proposition 15.59, the diffusion matrix, Dkℓ associated with the diffusion kernel Jk,ℓ (r, q, s) is spatially homogeneous, i.e., assume (8.33), and that the divergence of the diffusion matrix equals 0 at (0, t) ∀t, i.e., d
k=1
84
˜ kℓ (0, t) ≡ 0 ∀ℓ. (∂k D)
(15.195)
Our definition is equivalent to Kunita’s definition of the Stratonovich integral (cf. Kunita, loc.cit. Chap. 3.2, and our Chap. 7).
406
15 Appendix
Then t 0
J (r (s), q, s)w(dq, ◦ds) =
t
J (r (s), q, s)w(dq, ds),
(15.196)
0
i.e., the Stratonovich and the Itˆo integrals coincide in this particular case. Proof. 85 (i) Set σn,i j (r, t) :=
Ji j (r, q, s)φn (q)dq.
(15.197)
By Proposition 15.59 σn,i j (r (·), ·) are continuous square integrable semimartingales. Therefore, by (15.190) t t t σn,i j (r (s), s) ◦ β nj (ds) = σn,i j (r (s), s)β nj (ds) + σn,i j (r (t), t), β nj (ds) 0
0
t
0
and all we need is to compute the correction term [σn,i j (r (t), t), 0 β nj (ds)]. To this end, we must find the semimartingale representation of σn,i j (r (t)). By Itˆo’s formula t ∂ σn,i j (r (t), t) = σn,i j (r (0), 0) + σn,i j (r (s), s)ds ∂s 0 t d 1 2 ∂kℓ σn,i j (r (s), s)[rk , rℓ ](ds), + (∇σn,i j )(r (s), s) · r (ds) + 2 0 kℓ=1
where the assumptions allow us to interchange differentiation and integration. By (15.93) the processes of bounded variation do not contribute to the calculation t of [σn,i j (r (t), t), 0 β nj (ds)]. Hence, $
σn,i j (r (t), t),
≡
d
k=1
$
0
0
t
β nj (ds)
%
≡
t
(∂k σn,i j )(r (s), s)
$ d
& k=1
d
m ℓ=1
t
(∂k σn,i j )(r (s), s)rk (ds),
0
t
0
β nj (ds)
%⎫ ⎪ ⎪ ⎪ ⎪ ⎬
% t m n σm,kℓ (r (s), s)βℓ (ds) , β j (ds) . 0
⎪ ⎪ ⎪ ⎪ ⎭
(15.198)
Observe that β nj and βℓm are independent if (n, j) != (m, ℓ). Employing (15.151), we obtain t t ∂ n Ji j (r (t), q, t)φn (q)dq, β j (t) ≡ (∂k σn,i j (r (s), s))σn,k j (r (s), s)ds. 0 0 ∂rk (15.199) 85
The proof is adopted from Kotelenez (2007).
15.2 Stochastics
407
Adding up d
n
j=1
σn,i j (r, t)σn,k j (q, t) =
(J (r, p, t)J (q, p, t))ik d p.
Hence, summing up the correction terms, ⎞ ⎛ d d
∂ ⎝ σn,i j (r, t)σn,k j (q, t)q=r ⎠ ∂rk n k=1
j=1
d
= However,
Hence,
k=1
∂ ∂rk
(J (r, p, t)J (q, p, t))ik d p)q=r .
(J (r, p, t)J (q, p, t))ik d p = Dik (r − q, t).
d d
d
∂ (∂k σn,i j )(r, t)σn,k j (q, t)q=r ≡ Dik (r − q, t)q=r ≡ 0 ∂r k n k=1
k=1
j=1
(15.200) by assumption (15.198). We conclude that the sum of all correction terms equals 0, which implies (15.199). ⊔ ⊓ Lemma 15.61. Let ϕ ∈ Cc3 (Rd ; R) and assume the conditions and notation of Proposition 15.60. Then
t (∇ϕ)(r (s)) · (◦ σn (r (s), s)β n (ds)) 0
n
=
n
t
0
(15.201)
n
(∇ϕ)(r (s)) · (σn (r (s), s) ◦ β (ds)).
Proof. (i) As in the proof of Proposition 15.60, we first analyze the onedimensional t ˜ s)ds is coordinates of (15.201). Since the deterministic integral 0 F(r (s), Y, the same for Itˆo and Stratonovich integrals, we may in what follows assume, without loss of generality, F ≡ 0.
408
15 Appendix
(ii) By (15.187), d
t
(∂k ϕ)(r (s)) ◦ σn,k j (r (s), s)β nj (ds)
j=1 0 d t
=
(∂k ϕ)(r (s))σn,k j (r (s), s)β nj (ds)
⎫ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎬
⎪ ⎪ ⎤ ⎪ ⎪ ⎪ d ⎪ t
⎪ 1⎣ ⎪ n ⎦ ⎪ + σn,k j (r (s), s)β j (ds) . ⎪ (∂k ϕ)(r (s)), ⎪ ⎭ 2 0 0 j=1 ⎡
(15.202)
j=1
To compute the correction term we apply Itˆo’s formula to (∂k ϕ)(r (t)). Recalling that, by (15.93), only the martingale part in Itˆo’s formula contributes to the mutual quadratic variation, we obtain ⎤ ⎡ d t
1⎣ σn,k j (R(s), s)β nj (ds)⎦ ((∂k ϕ)(r (t)), 2 0 j=1 #& $ " d d t
1 2 m = (∂kℓ ϕ)(r (s)) σm,ℓi (r (s), s)βi (ds) 2 0 m ℓ=1 ℓ=1 % d t
n × σn,k j (r (s), s)β j (ds) j=1 0
=
" d 1 t
2
0
i.e.,
ℓ=1
2 (∂kℓ ϕ)(r (s))
#⎧ d ⎨
⎩
σn,ℓj (r (s), s)σn,k j (r (s), s)ds
j=1
⎤ ⎡ ⎫ d t ⎪
1⎣ ⎪ ⎪ ⎪ σn,k j (r (s), s)β nj (ds)⎦ (∂k ϕ)(r (t)), ⎪ ⎬ 2 0 j=1 d ⎪ ⎪ 1 t 2 ⎪ = (∂kℓ ϕ)(r (s))σn,ℓj (r (s), s)σn,k j (r (s), s)ds. ⎪ ⎪ ⎭ 2 0
⎫ ⎬ ⎭
.
(15.203)
j,ℓ=1
Summing up over all n yields: ⎤⎫ ⎡ d t ⎪
⎪ 1⎣ ∂ ⎪ ϕ (r (t)), σn,k j (r (s), s)β nj (ds)⎦ ⎪ ⎪ ⎬ 2 ∂rk n j=1 0 d ⎪ ⎪ 1 t 2 ⎪ ⎪ = (∂kℓ ϕ)(r (s))Dℓk (0, s)ds. ⎪ ⎭ 2 0
(15.204)
ℓ=1
Summing up (15.202) over all n and changing the notation for the indices we obtain from the previous formulas
15.2 Stochastics
409 t
0
=
(∂k ϕ)(r (s)) ◦
t
d
(∂k ϕ)(r (s))
0 d
1
+ 2
σn,kℓ (r (s), s)βℓn (ds)
n ℓ=1
d
Jkℓ (r (s), p, s)wℓ (d p, ds)
ℓ=1
t
ℓ=1 0
2 (∂kℓ ϕ)(r (s))Dℓk (0, s)ds.
(iii) Again by (15.187), d
2
(∂k ϕ)(r (t))σn,k j (r (t), t),
0
j=1
⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎭ ⎫ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎬
t
(∂k ϕ)(r (s))σn,k j (r (s), s) ◦ β nj (ds) 0 j=1 d t
(∂k ϕ)(r (s))σn,k j (r (s), s)β nj (ds) = 0 j=1 t d 1
+
⎫ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎬
⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ n β j (ds) . ⎪ ⎪ ⎪ ⎭
(15.205)
(15.206)
We now must compute the martingale term of (∂k ϕ)(r (t))σn,k j (r (t), t), using Itˆo’s formula. We verify that this term equals t σn,k j (r (s), s)(∇∂k ϕ)(r (s)) · J (r (s), p, s)w(d p, ds) 0 t + (∂k ϕ)(r (s))(∇σn,k j (r (s), s)) · J (r (s), p, s)w(d p, ds). 0
Employing the series representation (4.15) with the notation from (15.197), we obtain for the correction term t d 1
n (∂k ϕ)(r (t))σn,k j (r (t), t), β j (ds) 2 0 j=1 " d # d t
1
2 σn,k j (r (s), s) ∂kℓ ϕ (r (s)) = 2 j=1 0 ℓ=1 & d t
m × σm,ℓi (r (s), s)βi (ds) , β nj (ds) m i=1
+
0
d d t
1
(∂k ϕ)(r (s))(∂ℓ σn,k j )(r (s), s)) 2 j=1 0 ℓ=1 & d t
m × σm,ℓi (r (s), s)βi (ds) , β nj (ds) . m i=1
0
410
15 Appendix
Observing the orthogonality of the Brownian motions the correction term simplifies t d 1
n (∂k ϕ)(r (t))σn,k j (r (t), t), β j (ds) 2 0 j=1 $ " d % # t d t
1
n n 2 = σn,k j (r (s), s) β j (ds) ∂kℓ ϕ)(r (s))σn,ℓj (r (s), s β j (ds), 2 0 0 j=1 ℓ=1 " d # d t
1
+ (∂k ϕ)(r (s)) (∂ℓ σn,k j )(r (s), s))σn,ℓj (r (s), s)β nj (ds), 2 0 ℓ=1
j=1
×
t
0
β nj (ds)
.
whence by (15.151) the correction term satisfies the following equation: ⎫ t d ⎪ 1
⎪ n ⎪ ⎪ (∂k ϕ)(r (t))σn,k j (r (t), t), β j (ds) ⎪ ⎪ 2 ⎪ 0 ⎪ j=1 ⎪ ⎪ ⎪ d ⎬ t
1 2 = (∂kℓ ϕ)(r (s))σn,k j (r (s), s)σn,ℓj (r (s), s)ds (15.207) 2 ⎪ ⎪ ℓ, j=1 0 ⎪ ⎪ ⎪ d ⎪ ⎪ 1 t ⎪ ⎪ (∂k ϕ)(r (s))(∂ℓ σn,k j )(r (s), s))σn,ℓj (r (s), s)ds. ⎪ + ⎪ ⎭ 2 0 ℓ, j=1
It remains to sum up the correction terms over all n. First, we compute the sum of the second terms. Basically repeating the calculations of (15.200), ⎫ d
⎪ 1 t ⎪ ⎪ ((∂k ϕ)(r (s)) (∂ℓ σn,k j )(r (s), s))σn,ℓj (r (s), s) ds ⎪ ⎪ 2 ⎪ 0 ⎪ n n ℓ, j=1 ⎪ ⎪ ⎪ d ⎪ t
⎪ ∂ 1 ⎬ (∂k ϕ)(r (s)) Jk j (r, p, s)Jℓj (q, p, s)d pr =q=r (s) ds = 2 ∂r ℓ ⎪ ℓ, j=1 0 ⎪ ⎪ d ⎪ t ⎪ ∂ 1
⎪ ⎪ ⎪ = Dk,ℓ (r − q, s)r =q=r (s) ds (∂k ϕ)(r (s)) ⎪ ⎪ 2 ∂rℓ ⎪ 0 ⎪ ℓ=1 ⎭ = 0 by assumption (15.195). (15.208) Summing up over the first terms in the righthand side of (15.207) yields ⎫ d
⎪ 1 t 2 ⎪ (∂kℓ ϕ)(r (s)) σn,k j (r (s), s)σn,ℓj (r (s), s)ds ⎪ ⎪ ⎪ ⎪ 2 ⎬ n ℓ, j=1 0 (15.209) ⎪ d ⎪ ⎪ 1 t 2 ⎪ ⎪ (∂kℓ ϕ)(r (s))Dkℓ (0, s)ds. = ⎪ ⎭ 2 ℓ
0
15.2 Stochastics
411
Since the diffusion matrix Dkℓ is symmetric, we obtain by (15.204) that the correction terms for both sides of (15.201) are equal. ⊓ ⊔ We conclude this subsection with the following important observation:86 Theorem 15.62. Suppose that the conditions of Proposition 15.60 hold and let ϕ ∈ Cc3 (Rd ; R). Denoting by r (·) the solution of the (Iˆo) SODE (15.190) the following holds: ⎫ d ⎪ 1 2 ⎪ ⎪ ϕ(r (t)) ≡ (∂kℓ ϕ)(r (s))Dkℓ (r (s), s)ds ⎪ ⎪ ⎪ 2 ⎪ ⎪ k,ℓ=1 ⎪ ⎪ d ⎪ t
⎪ ⎪ ⎪ (∇ϕ(r (s)) · Jk,ℓ (r (s), q, s)wℓ (dq, ds) ⎪ + ⎪ ⎪ ⎪ 0 ⎬ k=1 t (15.210) ˜ + (∇ϕ)(r (s)) · F(r (s), Y(s), s)ds ⎪ ⎪ ⎪ 0 ⎪ ⎪ d t ⎪
⎪ ⎪ = (∇ϕ)(r (s)) · Jk,ℓ (r (s), q, s)wℓ (dq, ◦ds) ⎪ ⎪ ⎪ ⎪ 0 ⎪ k=1 ⎪ t ⎪ ⎪ ⎪ ⎪ ˜ ⎭ + (∇ϕ)(r (s)) · F(r (s), Y(s), s)ds. 0
Proof. The Itˆo formula (Theorem 15.50) yields the the first part of (15.210). The chain rule for Stratonovich integrals (Theorem 15.58) in addition to Lemma 15.61 implies that ϕ(r (t)) equals the righthand side of (15.210). ⊔ ⊓
15.2.7 MarkovDiffusion Processes We review the definition of Markovdiffusion processes and prove the Markov property for a class of semilinar stochastic evolution equations in Hilbert space. We then derive Kolmogorov’s backward and forward equations for Rd valued Markovdiffusion processes. The latter is also called the “Fokker–Planck equation.” It follows that the SPDEs considered in this volume are stochastic FokkerPlanck equations for (generalized) densities (or “number densities”). The general description and analysis of Markov processes is found in Dynkin (1965). We will restrict ourselves only to some minimal properties of continuous (nonterminating) Markovdiffusion processes (i.e., where the sample paths are a.s. continuous) with values in Rd . However, to allow applications to SPDEs and other infinite dimensional stochastic differential equations, we will initially consider continuous nonterminating processes with values in a separable metric space B. For the basic definitions we choose the “canonical” probability space  := C([0, ∞); B), endowed with the Borel σ algebra F := BC([0,∞);B) . 86
Cf. Theorem 5.3 in Kotelenez (2007).
412
15 Appendix
The first object is a family of transition probabilities87 P(s, ξ, t, B), 0 ≤ s ≤ t < ∞, ξ ∈ B, B ∈ BB (the Borel sets on B). By assumption, this family satisfies the following properties: Definition 15.63. (I) A family of probability measures on BB , P(s, ξ, t, ·), 0 ≤ s ≤ t < ∞, ξ ∈ B, is called a family of transition probabilities if (i) P(s, ·, t, B) is BB measurable for all 0 ≤ s ≤ t, B ∈ BB such that P(s, ·, s, B) = 1 B (·) for all s ≥ 0, (ii) and if for 0 ≤ s < u < t, ξ ∈ B and B ∈ BB the “ChapmanKolmogorov equation” holds: P(s, ξ, t, B) = P(u, η, t, B)P(s, ξ, u, dη). (15.211) B
(II) Suppose, in addition to the family of transition probabilities, there is a probability measure µ on BB . A probability measure P on (, BC([0,∞);B) ) is called a “Markov process” with transition probabilities P(s, ξ, t, B) and initial distribution µ if (15.212) P(ωB (0) ∈ B) = µ(B) ∀B ∈ BB . and if for all 0 ≤ s < u < t and B ∈ BB P(ωB (t) ∈ Bσ {ωB (u), 0 ≤ u ≤ s}) = P(s, ωB (s), t, B) a.s.,
(15.213)
where the quantity in the lefthand side of (15.213) is the conditional probability, conditioned on the “past” up to time s and ωB (·) ∈ . If P(s, ξ, t, ·) ≡ Ph (t − s, ξ, ·), where Ph (t − s, ξ, ·) is some timehomogeneous family of probability measures with similar properties as P(s, ξ, t, ·), then the corresponding Markov process is said to be “time homogeneous.” ⊔ ⊓ Property (15.213) is called the “Markov property.” It signifies that knowing the state of the system at time s is sufficient to determine the future. Obviously, a Markov process is not a stochastic process with a fixed distribution on the sample path but rather a family of processes, indexed by the initial distributions and having the same transition probabilities. It is similar to an ODE or SODE without specified initial value, where the “forcing function” describes the transition from a given state to the next state in an infinitesimal timestep. This begs the question of whether there is a deeper relation between SODEs (or ODEs)88 and Markov processes. Indeed, it has been Itˆo’s program to represent a large class of Markov processes by solutions of SODEs. To better understand this relation, let us look at a fairly general example of infinite dimensional stochastic differential equations related to semilinear SPDEs. 87 88
We follow the definition provided by Stroock and Varadhan (1979), Chap. 2.2. Notice that ODEs can be considered special SODEs, namely where the diffusion coefficient equals 0.
15.2 Stochastics
413
Let H and K be a real separable Hilbert spaces with scalar products &·, ·'H , &·, ·'K and norms · H , · K . Suppose there is a strongly continuous twoparameter semigroup U (t, s) on H.89 In addition to U (t, s), let B(s, ·) be a family of (possibly) nonlinear operators on H, Z ∈ H, s ≥ 0 such that U (t, s)B(s, ·) is bounded on H and measurable in (s, Z ) ∈ [0, t] × H for all t ≥ 0. Further, let W (·) be a Kvalued Brownian motion and C(s, ·) a family of (possibly) nonlinear operators on H with values in the space of linear operators from K into H. Suppose that (s, Z ) %→ U (t, s)C(s, Z ) is a jointly measurable map from [0, t] × H into the bounded linear operators from K into H for all t ≥ 0 such that for all Z ∈ H and t ≥0 t
E
0
U (t, s)C(s, Z )W (ds)2H < ∞.90
Consider the following stochastic evolution equation on H: t t X (t) = X s + U (t, u)B(u, X (u))du + U (t, u)C(u, X (u))W (du). s
s
(15.214) where X s is an adapted square integrable initial condition. We assume in what follows that (15.217) has a unique solution X (·, s, X s ), which is adapted, square integrable and such that for any two adapted square integrable conditions X s and Ys the following continuous dependence on the initial conditions holds sup EX (t, s, X s ) − X (t, s, Ys )2H ≤ ct¯ EX s − Ys 2H ∀t¯ < ∞,
s≤t≤t¯
(15.215)
where c(t¯) < ∞. If there is a generator A(t) of U (t, s) the solution of (15.214) is called a “mild solution.” We adopt this term, whether or not there is a generator A(t). Recall that in the finite dimensional setting of (4.10), as a consequence of global Lipschitz assumptions, (4.17) provides an estimate as in (15.215) if we replace the input processes Y˜ by the empirical processes with N coordinates. Similar global (or even local) Lipschitz assumptions and linear growth assumptions on (15.214) guarantee the validity of estimate (15.215) (cf. Arnold et al. (1980) or DaPrato and Zabczyk (1992)). Proposition 15.64. Under the preceding assumption the solution of (15.214) is a Markov process with transition probabilities P(s, Z , t, B) := E1 B (X (t, s, Z )), Z ∈ H, B ∈ BH . 89 90
(15.216)
Cf. Definition 15.41. t This condition implies that 0 U (t, s)C(s, Z )W (ds) is a regular Hvalued Gaussian random variable, i.e., it is a generalization an Hvalued regular Brownian motion whose covariance depends on t. Arnold et al. (loc.cit.) actually assumed W (·) to be regular on K, etc. However, DaPrato and Zabczyk (1992) provide many examples where W (·) can be cylindrical and the “smoothing effect” of the semigroup “regularizes” the Brownian motion.
414
15 Appendix
Proof. 91 (i) Let Gst := σ {W (v) − W (u) : s ≤ u ≤ v ≤ t} and G¯st the completed σ algebra. The existence of a solution X (t, s, Z , ω) of (15.214), which is measurable in (s, Z , ω) ∈ [0, t] × H ×  with respect to the σ algebra B[0,t] ⊗ BH ⊗ G¯st follows verbatim as the derivation of property 3) of Theorem 4.5.92 Hence the lefthand side of (15.216) defines a family of probability measures that is measurable in Z . (ii) Next, we show the Markov property (15.213) with X (·) replacing ωB (·). The following function f B (Z , ω) := 1 B (X (t, s, Z , ω)
is bounded and BH ⊗ G¯st measurable. Consider functions g(Z , ω) :=
n
h i (Z )φi (ω)
i=1
where all h i (·) are bounded BH measurable and all φi (·) are bounded G¯st measurable. Note that φi (·) are independent of σ {X (u), 0 ≤ u ≤ s}, whence E[φi (ω)σ {X (u), 0 ≤ u ≤ s}] = Eφi (·) = E[φi (ω)σ {X (s)}]. Therefore, E[g(X (s), ω)σ {X (u), 0 ≤ u ≤ s} = = E[g(X (s), ω)σ {X (s)}].
n
i=1
h i (X (s))E[φi (ω)σ {X (s)}]
(15.217) A variant of the monotone class theorem (cf. Dynkin (1961), Sect. 1.1, Lemma 1.2) implies that (15.217) can be extended to all bounded BH ⊗ G¯st measurable functions. Consequently, it holds for f B (Z , ω) and the Markov property (15.213) follows. (iii) For 0 ≤ s ≤ u ≤ t < ∞, by the uniqueness of the solutions of (15.214), X (t, s, Z ) = X (t, u, X (u, s, Z )) a.s. Thus, ⎫ P(s, Z , t, B) = E1 B (X (t, s, Z )) = E1 B (X (t, u, X (u, s, Z ))) ⎪ ⎬ = E[E[1 B (X (t, u, X (u, s, Z )))σ {X (u, s, Z )}]] = E[P(u, X (u, s, Z ), t, B)] ⎪ ⎭ = P(u, Z˜ , t, B)P(s, Z , u, d Z˜ ). H
(15.218) This is the Chapman–Kolmogorov equation (15.211) and the proof of Proposition 15.64 is complete. ⊔ ⊓ 91 92
We follow the proof provided by Arnold et al. (1980), which itself is a generalization of a proof by Dynkin (1965), Vol. I, Chap. 11, for finitedimensional SODEs. Cf. Chap. 6.
15.2 Stochastics
415
Corollary 15.65. The unique solution (r 1 (·, s, q 1 ), . . . , r N (·, s, q N ) of (4.10) is an R N d valued continuous Markov process. ⊔ ⊓ In what follows, we restrict ourselves to Markovdiffusion processes with values in Rd . Accordingly, the canonical probability space is (, F) := (C([0, ∞); Rd ), BC([0,∞);Rd ) ). We now write r (·) for an element of C([0, ∞); Rd ). Example 15.66. (I) P(s, q, t, B) :=
d
1 B (r )(2π(t − s))− 2 exp ( −
r − q2 )dr, 2(t − s)
(15.219)
which is the transition function of an Rd valued standard Brownian motion β(·). The most frequently used initial distribution is µ := δ0 . A Markov process, generated by the transition function (15.219), is best represented by the solution of the following (simple!) SODE dr = dβ r0 = q,
(15.220)
(II) Consider the ODE d r = F(r, t), r (s) = q. (15.221) dt Suppose the forcing function F is nice so that the ODE has a unique solution for all initial values (s, q), which is measurable in the initial conditions (s, q). Denote this solution r (t, s, q). Set P(s, q, t, B) := 1 B (r (t, s, q)).
(15.222)
(III) A more general example of a Markov process are the solutions of (4.10), considered in Rd N .93 In this case, we set P(s, q, t, B) := E1 B (r N (t, s, q)).
(15.223) ⊔ ⊓
Let P(s, q, t, B) be a family transition probabilities P(s, q, t, B) satisfies the requirements of Definition 15.62. From the Chapman–Kolmogorov equation we obtain a twoparameter semigroup U ′ (t, s) on the space of probability measures through (U ′ (t, s)µ)(·) =
P(s, q, t, ·)µ(dq).
(15.224)
Suppose that U ′ (t, s) is strongly continuous on the space of probability measures, endowed with the restriction of the norm of the total variation of a finite signed measures. By the Riesz representation theorem94 the transition probabilities also define in a canonical way a strongly continuous twoparameter semigroup U (t, s) of bounded operators on the space C0 (Rd ; R), setting for f ∈ C0 (Rd ; R) 93 94
Cf. Corollary 15.65. Cf., e.g., Folland (1984), Chap. 7.3.
416
15 Appendix
(U (t, s) f )(q) :=
P(s, q, t, dq) ˜ f (q). ˜
(15.225)
If the transition probabilities are timehomogeneous, the corresponding semigroups, U ′ (t, s) and U (t, s), depend only on t − s and have time independent generators, denoted A′ and A, respectively. To better understand the relation between Kolmogorov’s backward and forward equations and the general theory of semigroups let us first look at the timehomogeneous case. In this case we will write P(t, q, B) for the transition probabilities. C0 (Rd ; R) is the domain of the semigroup U (t), endowed with the supremum norm  · . If f ∈ Dom(A) we obtain 1 (A f )(q) := lim ((U (t) f )(q) − f (q)) = lim P(t, q, dr )( f (r ) − f (q)), t→0 t→0 t (15.226) where the limit is taken with respect to the norm  · . (15.226) is called “Kolmogorov’s backward equation” in weak form, where “backward” refers to the fact that the evaluation of the generator (for Markovdiffusions the closure of a differential operator) is computed with respect to the initial variable q.95 We have seen in Chap. 14, (14.38)–(14.39) that a timeinhomogeneous Markov process, represented by an SODE, can be transformed into a timehomogeneous Markov process by adding time as an additional dimension. However, the analysis of backward and forward equations in the original timeinhomogeneous setting provides more insight into the structure of the time evolution of the Markov process. Therefore, we will now derive the backward equation for the timeinhomogeneous case (with timedependent generator A(s)). Note that for the timehomogeneous case P(t − s, q, B) ∂ ∂ P(t − s, q, B)t−s=u = − P(t − s, q, B)t−s=u . ∂t ∂s Thus, we obtain in weak form the following “backward” equations for the twoparameter semigroups, U ′ (t, s) and U (t, s):96 −
∂ ′ ∂ U (t, s) = U ′ (t, s)A′ (s), − U (t, s) = U (t, s)A(s). ∂s ∂s
(15.227)
We denote the duality between measures and test functions by &·, ·'. Assuming f ∈ Dom(A(u)) ∀u, we then have ⎫ &(U′ (t, s − h) − U ′ (t, s))µ, f ' = &µ, (U (t, s − h) − U (t, s)) f ' ⎬ s s (15.228) = &µ, U (t, u)A(u) f 'du = &A′ (u)U ′ (t, u)µ, f 'du ⎭ s−h
95
96
s−h
The reader, familiar with the semigroups of operators on a Banach space and their generators, will recognize that the lefthand side of (15.226) is the definition of the generator of U (t) (cf., e.g., Pazy (1983)). For the analytic treatment of timedependent generators and the existence of associated twoparameter semigroups we refer the reader to Tanabe (1979), Chap. 5.2, where U (t, s) is called “fundamental solution.” Note that, in our setting, Tanabe considers twoparameter semigroups on the space of test functions with −A(t) being a generator.
15.2 Stochastics
417
We divide the quantities in (15.228) by h and pass to the limit. Choosing as the initial distribution at time s the Dirac measure δq yields “Kolmogorov’s backward equation” for the twoparameter flow of transition probabilities −
∂ P(s, q, t, ·) = (A′ (s)P(s, ·, t, ·))(q), ∂s
P(s, q, s, ·) = δq (·).
(15.229)
For a solution of an SODE (or ODE) we compute the generator A(s) of U (t, s) through Itˆo’s formula (or the simple chain rule). To simplify the notation, suppose that N = 1 in (4.10) and that the coefficients do not depend on the empirical process. Let f ∈ Cc2 (Rd ; R). Then (A(s) f )(q) := (F(q, s) · ∇ f )(q) ⎫ ⎧ d d ⎨
⎬
1 2 + f )(q), Jk j (q, q, ˜ s)Jℓj (q, q, ˜ s)dq˜ + σk⊥j (q, s)σℓj⊥ (q, s) (∂kℓ ⎭ ⎩ 2 k,ℓ=1
j=1
(15.230)
where σ ⊥ := σ1⊥ . The only term missing from Itˆo’s formula is the stochastic term, because it has mean 0 (15.230) and its generalizations and specializations are called the “ItˆoDynkin formula” for the generator. Expression (15.228) lead to the backward equation, by varying the backward time argument s. Let us present the analogous argument by varying the forward time t. Then, ⎫ &(U ′ (t + h, s) − U ′ (t, s))µ, f ' = &µ, (U (t + h, s) − U (t, s)) f ' ⎬ t+h t+h (15.231) = &µ, A(u)U (u, t) f &du = &A′ (u)U ′ (u, t)µ, f 'du. ⎭ t
t
We divide the quantities in (15.231) by h and pass to the limit. Choosing as the initial distribution at time s the Dirac measure δq yields “Kolmogorov’s forward equation” for the twoparameter flow of transition probabilities ∂ P(s, q, t, ·) = (A′ (t)P(s, q, t, ·)), ∂t
P(s, q, s, ·) = δq (·).
(15.232)
Next, suppose that P(s, q, t, B) has a density, p(s, q, t, r ), with respect to the Lebesgue measure so that 1 B P(s, q, t, dr ) = 1 B p(s, q, t, r )dr ∀t > s ≥ 0, q ∈ Rd , B ∈ B d . (15.233) Computing the duality of the righthand side of (15.227) with start in µ against a smooth test function f we obtain &(A′ (t) p(s, q, t, ·)µ(dq), f ' = µ(dq) p(s, q, t, r )(A(t) f )(r )dr. (15.234)
418
15 Appendix
We are mainly interested in the case when the generator A is the closure of a secondorder differential operator as in (15.230), which defines a Markovdiffusion process. If, for such an operator, the coefficients are sufficiently smooth, we can integrate by parts in (15.234) and obtain adjoint generator, A∗ , as an unbounded operator in H0 = L 2 (Rd , dr ) p(s, q, t, r )(A(t) f )(r )dr = (A∗ (t) p(s, q, t, ·))(r ) f (r )dr. (15.235) Put (15.232) and (15.235) together and suppose that the domain of A(t) is dense in C0 (Rd ; R) for all t ≥ 0. We then obtain “Kolmogorov’s forward equation” for the density or, as it is called in the physics and engineering literature, the “Fokker– Planck equation”: ∂ (15.236) p (s, q, t, r ) ≡ (A∗ p(s, q, t, ·))(r ), p(s, q, s, r ) = δq (r ). ∂t For the simplified version of (4.10), considered in (15.230), the Fokker–Planck equation is cast in the form ⎫ ∂ ⎪ ⎪ p (s, q, t, r ) = −(∇ · F(·, t) p(s, q, , t·))(r ) ⎪ ⎬ ∂t 3 d d
1 2 ⊥ ⎪ + Jk j (·, q, ˜ t)Jℓj (·, q, ˜ t)dq˜ + σk⊥j (·, t)σkℓ (·, t) (r ). ⎪ ∂kℓ p(s, q, t, ·) ⎪ ⎭ 2 k,ℓ=1
j=1
(15.237)
Remark 15.67. As we have mentioned previously, the independence of the coefficients of the empirical process and N = 1 was assumed for notational convenience. We now compare (15.237) with the SPDEs in (8.26) or, more generally, in (8.55). Replacing the deterministic density p(s, q, ·) by X (t, ·) and adding the stochastic terms, we conclude that the SPDEs derived in this book are stochastic Fokker– Planck equations for the (generalized) density X (t, ·). The generalized density corresponds to what is called “number density” in physics and engineering. ⊔ ⊓
15.2.8 MeasureValued Flows: Proof of Proposition 4.3 (i) The definition of µ(·) is equivalent to µ(t, dr ) = δϕ(t,q) (dr )µ0 (dq). So,
γ̟ (µ(t)) ≤
̟ (r )µ(t, dr ) =
̟ (ϕ(t, q))µ0 (dq).
15.2 Stochastics
419
Let β > 0 and consider ̟β (r ) := ̟ β (r ). By Itˆo’s formula ∀q t ̟β (ϕ(t, q))= ̟β (q) + (triangledown̟β )(ϕ(s, q)) · [db(s, q) + dm(s, q)] 0
d 1 2 + (∂kℓ ̟β )(ϕ(s, q))d[m k (s, q), m ℓ (s, q)]. 2 k,ℓ=1
Hence, by the BurkholderDaviesGundy inequality and the usual inequality for increasing processes, ⎫ E sup ̟β (ϕ(t, q) ⎪ ⎪ ⎪ 0≤t≤T ⎪ ⎪ T
⎪ d ⎪ ⎪ ⎪ ⎪ ≤ ̟β (q) + c E∂k ̟β (ϕ(s, q))dbk (s, q) ⎪ ⎪ ⎪ 0 k=1 ⎪ ⎪ 9 ⎪ ⎪ d ⎪ T ⎪ +cE 0  ∂k ̟β (ϕ(s, q))∂ℓ ̟β (ϕ(s, q))d[m k (s, q), m ℓ (s, q)] ⎪ ⎪ ⎪ ⎪ ⎪ k,ℓ ⎪ ⎪ T
⎪ d ⎪ ⎪ ⎪ 2 ⎪ +cE ∂k,ℓ ̟β (ϕ(s, q))d[m k (s, q), m ℓ (s, q)] ⎪ ⎪ ⎪ 0 k,ℓ ⎪ ⎪ ⎪ ⎪ d T ⎪ ⎪ ⎪ E̟β (ϕ(s, q))dbk (s, q) ≤ ̟β (q) + c 0 ⎪ ⎪ ⎪ k=1 ⎪ 6 ⎪ ⎪ 7 T d ⎬
7 2 8 (15.238) ̟β (ϕ(s, q))d{[m k (s, q)] + [m ℓ (s, q)]} +cE ⎪ ⎪ 0 k,ℓ ⎪ ⎪ ⎪ T
⎪ d ⎪ ⎪ ⎪ ⎪ ̟β (ϕ(s, q))d{[m k (s, q)] + [m ℓ (s, q)]} +cE ⎪ ⎪ ⎪ 0 k,ℓ ⎪ ⎪ ⎪ ⎪ ⎪ (by (15.46)) ⎪ ⎪ ⎪ T
d ⎪ ⎪ ⎪ ⎪ E̟β (ϕ(s, q))dbk (s, q) ≤ ̟β (q) + c ⎪ ⎪ ⎪ 0 k=1 ⎪ ⎪ 9 ⎪ ⎪ d ⎪ T ⎪ 2 ⎪ ⎪ E̟β (ϕ(s, q))d{[m k (s, q)] + [m ℓ (s, q)]} +c 0 ⎪ ⎪ ⎪ k,ℓ ⎪ ⎪ ⎪ T
d ⎪ ⎪ ⎪ ⎪ +cE ̟β (ϕ(s, q))d{[m k (s, q)] + [m ℓ (s, q)]} ⎪ ⎭ 0 k,ℓ
From Itˆo’s formula and (15.46) we obtain for β˜ > 0 E̟β˜ (ϕ(t, q) ≤ ̟β˜ (q) + c +cE
0
T
d
k,ℓ
T 0
d
k=1
E̟β˜ (ϕ(s, q))dbk (s, q)
̟β2˜ (ϕ(s, q))d {[m k (s, q)] + [m ℓ (s, q)]} .
420
15 Appendix
Gronwall’s inequality97 implies E̟β˜ (ϕ(t, q) ≤ cT,b,[m] ̟β˜ (q),
(15.239)
˜ the bounds of the derivatives where the finite constant cT,b,[m] depends on T, β, of b and {[m k (s, q)] + [m ℓ (s, q)]}. We employ (15.239) (with β˜ := 2β) to estimate the square root in the last inequality of (15.238): 9 d T c 0 E̟β2 (s, q)d{[m k (s, q)] + [m ℓ (s, q)]} k,ℓ
≤c
9
(15.240)
d T 0
E̟β2 (s, q)cT,b,[m] ds]} ≤ cˆT,b,[m] ̟β (q)
k,ℓ
To estimate the remaining integrals in (15.238) we again use Gronwall’s inequality and obtain altogether E sup ̟β (ϕ(t, q)) ≤ c¯T,b,[m] ̟β (q),
(15.241)
0≤t≤T
where c¯T,b,[m] depends only upon T, β and upon the bounds of the characteristics of ϕ, but not upon ϕ itself. In what follows we choose β = 1. We may change the initial measure µ0 on F0 measurable set of arbitrary small probability such that ess sup ω
̟ (q)µ0 (dq) < ∞.
Therefore, E ̟ (q)µ0 (dq) < ∞. We may use conditional expectations instead of absolute ones in the preceding calculations98 and assume in the first step that µ0 is deterministic. Hence E sup ̟ (ϕ(t, q)) is integrable against the initial measure 0≤t≤T
µ0 , and we obtain E sup 0≤t≤T
̟ (ϕ(t, q))µ0 (dq) ≤
≤ cT,b E
E sup ̟ (ϕ(t, q))µ0 (dq) 0≤t≤T
̟ (q)µ0 (dq) < ∞.
(15.242)
As a result we have a.s. sup γ̟ (µ(t)) ≤ sup
0≤t≤T
≤ 97 98
Cf. Proposition 15.5. Cf. (15.238).
0≤t≤T
̟ (ϕ(t, q))µ0 (dq)
sup ̟ (ϕ(t, q))µ0 (dq) < ∞.
0≤t≤T
(15.243)
15.2 Stochastics
421
(ii) We show that µ(·) is continuous (a.s.) in M∞,̟ . Let t > s ≥ 0. Then, γ̟ (µ(t) − µ(s)) = sup [ f (ϕ(t, q))̟ (ϕ(t, q)) f L ,∞ ≤1 − f (ϕ(s, q))̟ (ϕ(s, q))]µ0 (dq) ≤ sup f (ϕ(t, q)) − f (ϕ(s, q))̟ (ϕ(t, q))]µ0 (dq) f L ,∞ ≤1 + sup f (ϕ(s, q))[̟ (ϕ(t, q)) − ̟ (ϕ(s, q))]µ0 (dq) f L ,∞ ≤1
=: I (t, s) + I I (t, s).
Since the Lipschitz constant for all f in the supremum is 1 I (t, s) ≤ ρ(ϕ(t, q) − ϕ(s, q))̟ (ϕ(t, q))µ0 (dq). Further, ϕ(·, q) is uniformly continuous in [0, T ] and ρ(ϕ(t, q) − ϕ(s, q)) ≤ 1. Therefore, we conclude by (15.243) and Lebesgue’s dominated convergence theorem that a.s. sup I (t, s) = 0. lim δ→0 0≤s 0
H0 := L 2 (Rd , dr ) (H0 , · 0 ) = (W0,2,1 , · 0,2,1 )  182 Hm := Wm,2,1 . Hw  (4.31), (15.74) Id is the identity matrix in Rd I An (t)  occupation measure I˜An (t)  (3.7) n (t)  (3.9) I˜A,J n I˜An,⊥ (t)  (3.9) I˜n,⊥,c (t)  (3.15) A
(K, d K ) is some metric space with a norm d K , “l.s.t.” means “localizing stopping time” L p,loc (Rd × ; B)  Definition 6.2 L 0,Fs (K) is the space of Kvalued Fs measurable random variables L 0,F (C([s, T ]; K))  Kvalued continuous adapted and dt ⊗ d Pmeasurable processes L p,Fs (K)  those elements in L 0,Fs (K) with finite p−th moments of dK (ξ, e) L p,F (C([s, T ]; K)) are those elements in L 0,F (C([s, T ]; K)) such that p E sups≤t≤T dK (ξ(t), e) < ∞ L p,F (C([s, ∞); K))  space of processes ξ(·) such that ξ(·∧T )∈L p,F (C([s, T ]; K))∀T > s. L loc, p,F (C([s, T ]; K))  61 L loc, p,F (C((s, T ]; K))  61 L 1,F ([0, T ] × ) is the space of dt ⊗ dPmeasurable and adapted integrable processes L p,F ([0, T ]×; B) is the space of dt ⊗dPmeasurable and adapted p integrable Bvalued processes  Section 8.1
442
Symbols
Lk,ℓ,n (s) := ∂ n D˜ kℓ (s, r )r =0  (8.40) L(B)  the set of linear bounded operators from a Banach space B into itself m := {( p 1 , . . . , p m ) ∈ Rd·m : ∃i != j, i, j ∈ {1, . . . , m}, with pi = p j } (t, Dn , R, J n )  (3.1) m  mass of a small particle. mˆ  mass of a large particle. m(d x) := A(S d−1 )x d−1 d x, where A(S d−1 ) is the surface measure of the unit sphere in Rd m = m(m)  the smallest even integer with m > d + m + 2 Md×d  the d × d matrices N M f,s,d := {Xs : Xs = Xs,N := i=1 m i δrsi m+k := m + 1k m(d x)  (5.141) d M f  finite Borel measures on R M∞,̟ := {µ ∈ M∞ : ̟ (q)µ(dq) < ∞} µb :=µ, if γ̟ (µ) < b , and = γ̟µ(µ) b, if γ̟ (µ) ≥ b.  (9.2) ∧  denotes “minimum” ∨  denotes “maximum” n ∈ N  index characterizing the discretization n = (n 1 , . . . n d ) ∈ (N ∪ {0})d n := n 1 + . . . + n d . n ≤ m iff n i ≤ m i for i = 1, . . . , d. n < m iff n ≤ m and n < m.  ·  denoted both the absolute value and the Euclidean norm on Rd . · denotes the maximum norm on Rd . Fℓ  := supq Fℓ (q) denotes the supnorm. · L ,∞ denotes the bounded Lipschitz norm  (15.37)  f m := maxj≤m supr ∈Rd ∂ j f (r ) p f m, p,# := j≤m ∂ j f  p (r )#(r )dr, f m := f m,2,1  before (8.33) · f  dual norm on the finite Borel measures w.r.t. C0 (Rd , R) · L(B)  the norm of the set of linear bounded operators from a Banach space B into itself 0, i.e., the “null” element. O(d) are the orthogonal d × dmatrices over Rd ˆ d × Rd }N .  := {R (, F, Ft , P) is a stochastic basis with right continuous filtration 1 = (1, . . . , 1) 1k := (0, . . . , 0, 1, 0, . . . , 0) (πs,t f )(u) := f (u ∧ t), (u ≥ s). ˆ f −→ M f π2 (({r i (t)}∞ , X (t))) = X (t))  (8.69) π2 : M ˆ [0,T ]  η(·) %→ π N η(·) := η(g N (·))  (6.2) π N : M[0,T ] → M[0,T ],N ⊂ M
Symbols
443
d r n i , if for all i = 1, . . . , d n ≥ 0 and π (r ) = 0 otherwise. πn (r ) := %i=1 i n i
P(r ) :=
P ⊥ (r )
rr T r 2
 projection
:= Id − P(r )  projection ¯ a, B)  marginal transition probability distribution  (5.25) P(t, ψ(x) ∈ {log(x + e), 1(x)}  the set of measurable continuous semimartingale flows ϕ from Rd into itself  cf. Proposition 4.3 j ,..., jd ∂ j f = ∂r∂ j11...∂r jd f qn (t, λ, ι)  the position of a small particle with start in ( R¯ λ ] and velocity from Bι . ˆ d := Rd ∪ {⋄}. R ( R¯ λ ]  small cube, R¯ λ is the center of the cube rni (t)  position of ithlarge particle ∞ γ −1 −µu 1 e T (Du)du  the resolvent of D△ (µ − D2 △)−γ = Ŵ(γ ) 0 u γ
γ
Rµ := µγ (µ − 12 △)−γ , Rµ,D := µγ (µ − D2 △)−γ rˆ := (r 1 , r 2 ) ∈ R2d {r }  subspace spanned by r (Chap. 5) {r }⊥  subspace orthogonal to {r } (Section 5) ρ(r − q) := r − q ∧ 1  (4.1) ρ(r, ¯ q) = ρ(r − q), if r, q ∈ Rd and = 1, if r ∈ Rd and q = ⋄. ρ N (r N , q N ) := max1≤i≤N ρ(ri , qi )
(Sσ f )(r )  rotation operator  Chap. 10 S ′  Schwarz space of tempered distributions over Rd σ {·} denotes the σ −algebra generated by the the quantities in the braces [·, ·]  the mutual quadratic variation of square integrable martingales. '  summation, restricted to a set Cn  after (3.20) A τn is a sequence of localizing stopping times τ (c, b, ω) := infn τn (c, b, ω)  stopping time  (9.8)
Uh = shift operator  Chap. 10 ̟ (r ) = (1 + r 2 )−γ  weight function ̟β (r ) := ̟ β (r ) (β > 0) w¯  average speed of small particles wℓ (dr, dt) be i.i.d. real valued spacetime white noises on Rd ×R+ , ℓ = 1, . . . , d w(dp, ˇ t) := w(dp, T − t) − w(dp, T ) Wm, p,# := { f ∈ Bd,1,m : f m, p,# < ∞}.
444
Symbols
X N (t)  empirical (measure) process for N large particles, associated with (4.10). X˜N (t)  (3.7) X (t)  (9.55) 0 ∈ Md×d  the matrix with all entries being equal to 0
References
Adams, R.A. (1975), Sobolev Spaces. Academic Press, New York. Akhiezer, N.I. and Glasman, I.M. (1950), Theory of Linear Operators in Hilbert space. State Publisher of TechnicalTheoretical Literature, Moscow, Leningrad (in Russian). Albeverio, S., Haba, Z. and Russo, F. (2001), A TwoSpace Dimensional Semilinear Heat Equations Perturbed by (Gaussian) White Noise. Probab. Th. Rel. Fields 121, 319–366. Albeverio, S. and R¨ockner, M. (1991), Stochastic Differential Equations in Infinite Dimensions: Solutions via Dirichlet Forms. Probab. Th. Rel. Fields 89, 347–386. Aldous, D.J. (1985), Exchangeability and Related Topics. Ecole d’Ete de Probabilites de SaintFlour XIII—1985, Lecture Notes in Math., 1117, Springer, Berlin Heidelberg New York, pp. 1–198. Arnold, L. (1973), Stochastische DifferentialgleichungenTheorie und Anwendungen. R. Oldenburg, M¨unchenWien (English translation: Stochastic differential equations and applications, Wiley, New York). Arnold, L., Curtain, R.F. and Kotelenez, P. (1980), Nonlinear Stochastic Evolution Equations in Hilbert Space. Universit¨at Bremen, Forschungsschwerpunkt Dynamische Systeme, Report # 17. Asakura, S. and Oosawa, F. (1954), On Interactions between Two Bodies Immersed in a Solution of Macromolecules. J. Chem. Phys. 22, 1255–1256. Balakrishnan, A.V. (1983), On Abstract Bilinear Equations with White Noise Inputs. Appl. Math. Optim. 10, 359–366. Baxendale, P. and Rozovsky, B. (1993), Kinematic Dynamo and Intermittence in a Turbulent Flow. Geophys. Astrophys. Fluid Dyn. 73, 33–60. Bauer, H. (1968), Wahrscheinlichkeitstheorie und Grundz˝uge der Maßtheorie. de Gruyter & Co., Berlin (in German). Bensoussan, A., Glowinski, R. and R˘as¸canu, A. (1992), Approximation of Some Stochastic Differential Equations by the Splitting Up Method. Appl. Math. Optim. 25, 81–106.
445
446
References
Bhatt, A.G., Kallianpur, G. and Karandikar, R.I. (1995), Uniqueness and Robustness of Solutions of MeasureValued Equations of Nonlinear Filtering. Ann. Probab. 23(4), 1895–1938. Bichteler, K. (2002), Stochastic Integration with Jumps. Encyclopedia of Mathematics and its Applications 89, Cambridge University Press, Cambridge, New York. Billingsley, P. (1968), Convergence of Probability Measures. Wiley, New York. Bl¨omker, D. (2000), Stochastic Partial Differential Equations and Surface Growth. Wißner Verlag, Augsburg. Blount, D. (1996), Diffusion Limits for a Nonlinear Density Dependent Space–Time Population Model. Ann. Probab. 24(2), 639–659. Bojdecki, T. and Gorostiza, L.G. (1986), Langevin Equations for S ′ Valued Gaussian Processes and Fluctuation Limits of Infinite Particle Systems. Probab. Th. Rel. Fields 73, 227–244. Bojdecki, T. and Gorostiza, L.G. (1991), Gaussian and NonGaussian Distribution Valued OrnsteinUhlenbeck Processes. Can. J. Math. 43(6), 1136–1149. Bogachev, V.I. (1997), Gaussian Measures. Nauka, Moscow (in Russian). Borkar, V.S. (1984), Evolution of Interacting Particles in a Brownian Medium. Stochastics 14, 33–79. ´ ements de Math´ematique – Livre IV: Int´egration. Nauka, Bourbaki, N. (1977), El´ Moscow (Russian Translation). Brze´zniak, Z. and Li, Y. (2006), Asymptotic Compactness and Absorbing Sets for 2D Stochastic NavierStokes Equations on Some Unbounded Domains. Transactions of the American Mathematical Society S00029947(06)039237. Carmona, R. and Nualart, D. (1988a), Random Nonlinear Wave Equations: Smoothness of the Solutions. Probab. Th. Rel. Fields 79, 469–508. Carmona, R. and Nualart, D. (1988b), Random Nonlinear Wave Equations: Propagation of Singularities. Ann. Probab. 16(2), 730–751. Carmona, R. and Rozovsky, B. (1999), Stochastic Partial Differential Equations: Six Perspectives. Mathematical Surveys and Monographs, Vol. 64, American Mathematical Society. ChojnowskaMichalik, A. (1976), Stochastic Differential Equations in Hilbert Spaces. Ph.D. Thesis. Institute of Mathematics, Polish Academy of Science. Chorin, A.J. (1973), Numerical Study of a Slightly Viscous Flow. J. Fluid Mech. 57, 785–796. Chow, P.L. (1976), Function Space Differential Equations Associated with a Stochastic Partial Differential Equation. Indiana University Mathematical J. 25(7), 609–627. Chow, P.L. (1978), Stochastic Partial Differential Equations in Turbulence Related Problems. In: BharuchaReid, A.T. (ed.) Probability Analysis and Related Topics, Vol. I, Academic Press, New York, pp. 1–43. Chow, P.L. (2002), Stochastic Wave Equations with Polynomial Nonlinearity. Ann. Appl. Probab. 12(1), 361–381. Chow, P.L. and Jiang, J.L. (1994), Stochastic Partial Differential Equations in H¨older Spaces. Probab. Th. Rel. Fields 99, 1–27.
References
447
Coddington, E.A. and Levinson, N. (1955), Theory of Ordinary Differential Equations. McGrawHill, New York. Crauel, H., Debussche, A. and Flandoli, F. (1997), Random Attractors. J. Dyn. Differ. Eqns. 9(2), 307–341. Crisan, D. (2006), Particle Approximations for a Class of Stochastic Partial Differential Equations. Appl. Math. Optim. 54, 293–314. Curtain, R.F. (1981), Markov Processes Generated by Linear Stochastic Evolution Equations. Stochastics 5, 135–165. Curtain, R.F. and Falb, P.L. (1971), Stochastic Differential Equations in Hilbert Space. J. Differ. Eqns. 10, 412–430. Curtain, R.F. and Kotelenez, P. (1987), Stochastic Bilinear Spectral Systems. Stochastics 20, 3–15. Curtain, R.F. and Pritchard, A.J. (1978), Infinite Dimensional Linear Systems Theory. SpringerVerlag, Lecture Notes in Control and Information Sciences, Vol. 8, Berlin Heidelberg New York. Dalang, R. (1999), Extending Martingale Measure Stochastic Integral with Applications to Spatially Homogeneous SPDEs. EJP 4, 1–29. Dalang, R. and Mueller, C. (2003), Some Nonlinear SPDEs that are Second Order in Time. EJP 8(1), 1–21. Dalang, R., Mueller, C. and Tribe, R. (2006), A FeynmanKacType Formula for the Deterministic and Stochastic Wave Equations and Other SPDEs. Preprint. Dalang, R. and Nualart, D. (2004), Potential Theory for Hyperbolic SPDEs J. Funct. Anal. 227, 304–337. Dalang, R. and SanzSol, M. (2005), Regularity of the Sample Paths of a Class of Second Order SPDEs J. Funct. Anal. 227, 304–337. Dalang, R. and Walsh, J. (2002), TimeReversal in Hyperbolic SPDEs. Ann. Probab. 30(1), 213–252. Daletskii, Yu. L. and Goncharuk, N. Yu. (1994), On a Quasilinear Stochastic Differential Equation of Parabolic Type. Stochastic Anal. Appl. 12(1), 103–129. Da Prato, G. (1982), Regularity Results of a Convolution Stochastic Integral and Applications to Parabolic Stochastic Equations in a Hilbert Space. Conferenze del Seminario Matematico dell’Universtit´a di Bari, Nor. 182, Laterza. Da Prato, G. (2004), Kolmogorov Equations for Stochastic PDEs. Birkh¨auser, Boston Basel Berlin. Da Prato, G. and Debussche, A. (2002), 2D NavierStokes Equations Driven by a space–time White Noise. J. Funct. Anal., 196(1), 180–210. Da Prato, G. and Debussche, A. (2003), Strong Solutions to the Stochastic Quantization Equations. Ann. Probab. 31(4), 1900–1916. Da Prato, G., Iannelli, M. and Tubaro, L. (1982), Some Results on Linear Stochastic Differential Equations in Hilbert Spaces. Stochastics 23, 1–23. Da Prato, G., Kwapien, S. and Zabczyk, J. (1987), Regularity of Solutions of Linear Stochastic Equations in Hilbert Spaces. Stochastics, 23, 1–23. Da Prato, G. and Zabczyk, J. (1991), Smoothing Properties of the Kolmogorov Semigroups in Hilbert Space. Stochast. Stochast. Rep. 35, 63–77.
448
References
Da Prato, G. and Zabczyk, J. (1992), Stochastic Equations in Infinite Dimensions. Cambridge University Press, Cambridge. Da Prato, G. and Zabczyk, J. (1996), Ergodicity for Infinite Dimensional Systems. Cambridge University Press, London Mathematical Society, Lecture Note Series 229, Cambridge. Davies, E.B. (1980), OneParameter Semigroups. Academic Press, New York. Dawson, D.A. (1972), Stochastic Evolution Equations. Math. Biosci. 15, 287–316. Dawson, D.A. (1975), Stochastic Evolution Equations and Related Measure Processes. J. Multivariate Anal. 5, 1–52. Dawson, D.A. (1993), Private Communication. Dawson, D.A. (1993), MeasureValued Markov Processes. Ecole d’Ete de Probabilites de SaintFlour XXI—1991, Lecture Notes in Math., 1541, Springer, Berlin, pp. 1–260. Dawson, D.A. and Gorostiza, L. (1990), Generalized Solutions of a Class of Nuclear Space Valued Stochastic Evolution Equations. Appl. Math. Optim. 22, 241–263. Dawson, D.A. and Hochberg, K.L. (1979), The Carrying Dimension of a Stochastic Measure Diffusion. Ann. Probab. 7, 693–703. Dawson, D.A., Li, Z. and Wang, H. (2003), A Degenerate Stochastic Partial Differential Equation for the Putely Atomic Superprocess with Dependent Spatial Motion. Infinite Dimensional Analysis, Quantum Probability and Related Topics 6(4), 597–607. Dawson, D.A. and Vaillancourt, J. (1995), Stochastic McKeanVlasov Equations. No. DEA 2. 199–229. Dawson, D.A., Vaillancourt, J. and Wang, H. (2000). Stochastic Partial Differential Equations for a Class of Interacting MeasureValued Diffusions. Ann. Inst. Henri Poincare, Probabilites et Statistiques 36(2), 167–180. De Acosta, A. (1982), Invariance Principles in Probability for Triangular Arrays of BValued Random Vectors and Some Applications. Ann. Probab. 2, 346–373. Dellacherie, C. (1975), Capacit´es et processus stochastiques. Mir, Moscow (in Russian – Translation from French (1972), Springer, Berlin Heidelberg New York). Dieudonn´e, J. (1969), Foundations of Modern Analysis. Academic Press, New York. Doering, C.R. (1987), Nonlinear Parabolic Stochastic Differential Equations with Additive Colored Noise on Rd × R+ : A Regulated Stochastic Quantization. Commun. Math. Phys. 109, 537–561. DonatiMartin, C. and Pardoux, E. (1993), White Noise Driven SPDEs with Reflection. Probab. Th. Rel. Fields 95, 1–24. Donnelly, P. and Kurtz, T.G. (1999), Particle Representations for MeasureValued Population Models. Ann. Probab. 27, 166–205. Dorogovtsev, A. (2004a), Private Communication. Dorogovtsev, A. (2004b), One Brownian Stochastic Flow. Th. Stochast. Process. 10(26), 3–4, 21–25.
References
449
Dorogovtsev, A. and Kotelenez, P. (2006), Stationary Solutions of Quasilinear Stochastic Partial Differential Equations. Preprint, Department of Mathematics, CWRU. Duan, J., Lu, K. and Schmalfuss, B. (2003), Invariant Manifolds for Stochastic Partial Differential Equations. Ann. Probab. 31(4), 2109–2135. Duan, J., Lu, K. and Schmalfuss, B. (2004), Smooth Stable and Unstable Manifolds for Stochastic Evolutionary Equations. J. Dyn. and Differ. Eqns. 16(4), 949–972. Dudley, R. (1989), Real Analysis and Probability. Wadsworth and Brooks, Belmont, California. Dunford, N. and Schwartz, J. (1958), Linear Operators, Part I. Interscience, New York. D¨urr, D., Goldstein, S. and Lebowitz, J.L. (1981), A Mechanical Model of Brownian Motion. Commun. Math. Phys. 78, 507–530. D¨urr, D., Goldstein, S. and Lebowitz, J.L. (1983), A Mechanical Model for the Brownian Motion of a Convex Body. Z.Wahrscheinlichkeitstheorie verw. Gebiete. 62, 427–448. Dynkin, E.B. (1961), Die Grundlagen der Theorie der Markoffschen Prozesse. Springer Verlag, Berlin G¨ottingen Heidelberg. Dynkin, E.B. (1965), Markov Processes. Vol. I and II. Springer Verlag, Berlin G¨ottingen Heidelberg. Ehm, W., Gneiting, T. and Richards, D. (2004), Convolution Roots of Radial Positive Functions with Compact Support. Transactions of the American Mathematical Society 356, 4655–4685. ¨ Einstein, A. (1905), Uber die von der molekularkinetischen Theorie der W¨arme gefordete Bewegung von in ruhenden Fl¨ussigkeiten suspendierten Teilchen. Ann.d.Phys. 17 (quoted from the English translation: (1956) Investigation on the Theory of Brownian Movement), Dover Publications, New York. Erwe, F. (1968), Differential und Integralrechnung II. Bibliographisches Institut, Mannheim. Ethier, S.N. and Kurtz, T.G. (1986), Markov Processes – Characterization and Convergence. Wiley, New York. Faris, W.G. and JonaLasinio, G. (1982), Large Fluctuations for a Nonlinear Heat Equation with Noise. J. Phys. A. Math. Gen. 15, 3025–3055. Flandoli, F. (1995), Regularity Theory and Stochastic Flows for Parabolic SPDEs. Gordon and Breach, London. Flandoli, F. and Maslowski, B. (1995), Ergodicity of the 2D NavierStokes Equation Under Random Perturbations. Comm. Math. Phys. 172(1), 119–141. Fleming, W.H. and Viot, M. (1979), Some MeasureValued Markov Processes in Population Genetics Theory. Indian University Mathematics Journal, 28(5), 817–843. Folland, G.B. (1984), Real Analysis – Modern Techniques and their Applications. Wiley & Sons, New York. Fouque, J.P. (1994), Private Communication. Friedman, A. (1975), Stochastic Differential Equations and Applications, Vol. 1. Academic Press, New York San Francisco London.
450
References
Friedman, A. (1976), Stochastic differential equations and applications, Vol. 2. Academic Press, New York San Francisco London. Fukushima, M. (1980), Dirichlet Forms and Markov Processes. NorthHolland/ Kodansha. Amsterdam Oxford New York. Funaki, T. (1983), Random Motion of Strings and Related Stochastic Evolution Equations. Nagoya Math. J. 89, 129–193. Funaki, T. and Spohn, H. (1997), Motion by Mean Curvature from the GinzburgLandau ∇φ Interface Model. Commun. Math. Phys. 185, 1–36. G¨artner, J. (1988): On the McKeanVlasov limit for interacting diffusions. Math. Nachr. 137, 197–248. Gel’fand I.M. and Vilenkin, N.Ya. (1964), Generalized Functions, Vol 4. Academic Press, New York. Giacomin, G., Lebowitz, J.L. and Presutti, E. (1999), Deterministic and Stochastic Hydrodynamic Equations Arising from Simple Microscopic Model Systems. In: Stochastic Partial Differential Equations: Six Perspectives, Carmona R.A. and Rozovskii B. (eds.), Mathematical Surveys and Monographs, Vol. 64, American Mathematical Society, pp. 107–152. Gikhman, I.I. and Mestechkina, T.M. (1983), A Cauchy Problem for Stochastic First Order Partial Differential Equations and Their Applications. Preprint (in Russian). Gikhman, I.I. and Skorokhod, A.V. (1968), Stochastic Differential Equations. Naukova Dumka, Kiev (in Russian – English Translation (1972): Stochastic Differential Equations. Springer, Berlin Heidelberg New York). Gikhman, I.I. and Skorokhod, A.V. (1971), Theory of Random Processes. Vol. I. Nauka, Moscow (in Russian – English Translation (1974): The Theory of Stochastic Processes I. Springer, Berlin Heidelberg New York). Gikhman, I.I. and Skorokhod, A.V. (1982), Stochastic Differential Equations and Their Applications. Naukova Dumka, Kiev (in Russian). Goetzelmann, B., Evans, R. and Dietrich, S. (1998), Depletion Forces in Fluids. Phys. Rev. E 57(6), 6785–6800. Goncharuk, N. and Kotelenez, P. (1998), Fractional Step Method for Stochastic Evolution Equations. Stoch. Proc. Appl. 73, 1–45. Gorostiza, L.G. (1983), High Density Limit Theorems for Infinite Systems of Unscaled Branching Brownian Motions. Ann. Probab. 11(2), 374–392. Gorostiza, L.G. and Le´on (1990), Solutions of Stochastic Evolution Equations in Hilbert Space. Preprint. Greksch, W. and Tudor, C. (1995), Stochastic Evolution Equations. A Hilbert Space Approach. Akademie Verlag, Berlin. Gy¨ongy, I. (1982), On Stochastic Equations with Respect to Semimartingales III. Stochastics 7, 231–254. Gross, L. (1965), Abstract Wiener Spaces. Proc. 5th. Berkeley Sym. Math. Stat. Prob. 2, 31–42. Gurtin, M.E. (1981), An Introduction to Continuum Mechanics. Mathematics in Science and Engineering, Vol. 158, Academic Press, New York. Haken, H. (1983), Advanced Synergetics. Springer, Berlin Heidelberg New York.
References
451
Heisenberg, W. (1958), The Copenhagen Interpretation of Quantum Physics. In Physics and Philosophy, Volume Nineteen of World Perspectives, Harper and Row New York. Holden H., Øksendal, B., Ubøe, J. and Zhang, T. ((1996), Stochastic Partial Differential Equations. A Modeling, White Noise Functional Approach. Birkh¨auser, Boston. Holley, R. (1971), The Motion of a Heavy Particle in an Infinite OneDimensional Gas of Hard Spheres. Z. Wahrscheinlickeitstheor. Verw. Geb. 17, 181219. Holley, R. and Stroock, D.W. (1978), Generalized OrnsteinUhlenbeck Processes and Infinite Particle Branching Brownian Motions. Publ. RIMS, Kyoto Univ. 14, 741–788. Ibragimov, I.A. (1983), On Smoothness Conditions for Trajectories of Random Functions. Theory Probab. Appl. 28(2), 240–262. Ichikawa, A. (1982), Stability of Semilinear Stochastic Evolution Eqautions. J. Math. Anal. Appl. 90, 12–44. Ikeda, N. and Watanabe, S. (1981), Stochastic Differential Equations and Diffusion Processes. North Holland, New York. Iscoe, I. (1988), On the Supports of MeasureValued Critical Branching Brownian Motion. Ann. Probab. 16(1), 200–221. Iscoe, I. and McDonald, D. (1989), Large Deviations for ℓ2 Valued OrnsteinUhlenbeck Processes. Ann. Probab. 17, 58–73. I’lin, A.M. and Khasminskii, R.Z. (1964), On equations of Brownian motions. Probab. Th. Appl. IX(3), 466–491 (in Russian). Itˆo, K. (1944), Stochastic Integral. Proc. Japan. Acad. Tokyo, 20, 519–524. Itˆo, K. (1983), DistributionValued Processes Arising from Independent Brownian Motions. Math. Z 182, 17–33. Itˆo, K. (1984), Foundation of Stochastic Differential Equations in Infinite Dimensional Spaces. CBMSNSF Regional Conference Series, SIAM. Jacod, J. (1985), Theoremes limite pour les processus. Ecole d’Ete de Probabilites de SaintFlour XIII—1985, Lecture Notes in Math., 1117, Springer, Berlin Heidelberg New York, pp. 298–409 Jetschke, G. (1986), On the Equivalence of Different Approaches to Stochastic Partial Differential Equations. Math. Nachr. 128, 315–329. Kallianpur, G. and Wolpert, R. (1984), Infinite Dimensional Stochastic Diffrential Equation Models for Spatially Distributed Neurons. Appl. Math. Optim. 12, 125–172. Kallianpur, G. and Xiong, J. (1995), Stochastic Differential Equations in Infinite Dimensional Spaces. Institute of Mathematical Statistics, Lecture Notes – Monograph Series, Hayward, California. Kantorovich, L.V. and Akilov, G.P. (1977), Functional Analysis. Nauka, Moscow (in Russian). Kato, T. (1976), Perturbation Theory for Linear Operators. Springer, Berlin Heidelberg New York. Khasminskii, R.Z. (1969), Stability of Systems of Differential Equations Under Random Perturbations of their Parameters. Moscow, Nauka (in Russian).
452
References
Kim, K.H. (2005), L p Estimates for SPDE with Discontinuous Coefficients in Domains. EJP 10(1), 1–20. KnochePrevot, C. and R¨ockner, M. (2006), A Concise Course on Stochastic Partial Differential Equations. Lecture Notes, Preprint. Kolmogorov, A.N. (1956), On Skorokhod’s Convergence. Theory Probab. Appl. 1(2), 239–247 (in Russian). Konno, N. and Shiga, T (1988), Stochastic Differential Equations for Some Measure Valued Diffusions. Prob. Th. Rel. Fields 79, 201–225. Korotkov, V.B. (1983), Integral Operators. Nauka, Moscow (in Russian). Kotelenez, P. (1982), A Submartingale Type Inequality with Applications to Stochastic Evolution Equations. Stochastics 8. 139–151. Kotelenez, P. (1984), A Stopped Doob Inequality for Stochastic Convolution Integrals and Stochastic Evolution Equations. Stoch. Anal. Appl. 2(3), 245–265. Kotelenez, P. (1985), On the Semigroup Approach to Stochastic Evolution Equations. In Arnold, L. and Kotelenez, P. (eds.) Stochastic space–time Models and Limit Theorems, 95–139.D. Reidel, Dordrecht. Kotelenez, P. (1986), Law of Large Numbers and Central Limit Theorem for Linear Chemical Reactions with Diffusion. Probab. Ann. Probab. 14(1), 173–193. Kotelenez, P. (1987), A Maximal Inequality for Stochastic Convolution Integrals on Hilbert Space and space–time Regularity of Linear Stochastic Partial Differential Equations. Stochastics 21, 345–458. Kotelenez, P. (1988), High Density Limit Theorems for Nonlinear Reactions with Diffusion. Probab. Th. Rel. Fields 78, 11–37. Kotelenez, P. (1992a), Existence, Uniqueness and Smoothness for a Class of Function Valued Stochastic Partial Differential Equations. Stochast. Stochast. Rep. 41, 177–199. Kotelenez, P. (1992b), Comparison Methods for a Class of Function Valued Stochastic Partial Differential Equations. Probab. Th. Rel. Fields 93, 1–19. Kotelenez, P. (1995a), A Stochastic Navier Stokes Equation for the Vorticity of a Twodimensional Fluid. Ann. Appl. Probab. 5(4) 1126–1160. Kotelenez, P. (1995b), A Class of Quasilinear Stochastic Partial Differential Equations of McKeanVlasov Type with Mass Conservation. Probab. Th. Rel. Fields 102, 159–188 Kotelenez, P. (1995c), Particles, Vortex Dynamics and Stochastic Partial Differential Equations. In: Robert J. Adler et al. (eds.), Stochastic Modelling in Physical Oceanography. Birkh¨auser, Boston, pp. 271–294. Kotelenez, P. (1996), Stochastic Partial Differential Equations in the Construction of Random Fields from Particle Systems. Part I: Mass Conservation. CWRU, Department of Mathematics, Preprint, pp. 96–143. Kotelenez, P. (1999), Microscopic and Mesoscopic Models for Mass Distributions. In: Stochastic Dynamics, Crauel, H. and Gundlach, M. (eds.), (in honor of Ludwig Arnold), Springer, Berlin Heidelberg New York. Kotelenez, P. (2000), Smooth Solutions of Quasilinear Stochastic Partial Differential Equations of McKeanVlasov Type. In: Skorokhod’s Ideas in Probability Theory,
References
453
Korolyuk, V. Portenko, and N. Syta, H. (eds.), (in honor of A.V. Skorokhod), National Academy of Sciences of Ukraine, Institute of Mathematics, Kyiv. Kotelenez, P. (2002), Derivation of Brownian Motions from Deterministic Dynamics of Two Types of Particles. CWRU, Department of Mathematics, Technical Report No. 02–149 Kotelenez, P. (2005a), From Discrete Deterministic Dynamics to Stochastic Kinematics – A Derivation of Brownian Motions. Stochast. Dyn., 5(3), 343–384. Kotelenez, P. (2005b), Correlated Brownian Motions as an Approximation to Deterministic MeanField Dynamics. Ukrainian Math. J. T 57(6), 757–769. Kotelenez, P. (2007), Itˆo and Stratonovich Stochastic Partial Differential Equations. Transition from Microscopic to Macroscopic Equations. Quarterly of Applied Mathematics. (to appear). Kotelenez, P. and Kurtz, T.G. (2006), Macroscopic Limit for Stochastic Partial Differential Equations of McKeanVlasov Type. Preprint. Kotelenez, P., Leitman M. and Mann, J.A., Jr. (2007), On the Depletion Effect in Colloids. Preprint. Kotelenez, P. and Wang, K. (1994), Newtonian Particle Mechanics and Stochastic Partial Differential Equations. In: Dawson, D.A. (ed.), Measure Valued Processes, Stochastic Partial Differential Equations and Interacting Systems, Centre de Recherche Mathematiques, CRM Proceedings and Lecture Notes, Vol. 5, 130–149. Krylov, N.V. (1977), Controlled Diffusion Processes. Nauka, Moscow (in Russian). Krylov N.V. (1999), An Analytical Approach to SPDEs. In: Stochastic Partial Differential Equations: Six Perspectives, Carmona, R.A. and Rozovskii, B. (eds.), Mathematical Surveys and Monographs, Vol. 64, American Mathematical Society, pp. 185–242. Krylov, N.V. (2005), Private Communication. Krylov, N.V. and Rozovsky, B.L. (1979), On stochastic evolution equations. Itogi Nauki i tehniki, VINITI, 71–146 (in Russian). Kunita, H. (1990), Stochastic Flows and Stochastic Differential Equations. Cambridge University Press, Cambridge, New York. Kuo, H.H. (1975), Gaussian measures in Banach spaces. Springer, Berlin Heidelberg New York. Kupiainen, A. (2004), Statistical Theories of Turbulence. In: Random Media (J. Wehr ed.) Wydawnictwa ICM, Warszawa. Kurtz, T.G. (1975), Semigroups of Conditioned Shifts and Approximation of Markov Processes. Ann. Probab. 3, 618–642. Kurtz, T.G. and Protter, P.E. (1996), Weak Convergence of Stochastic Integrals and Differential Equations II: Infinite Dimensional Case. Probabilistic Models for Nonlinear Partial Differential Equations, Lecture Notes in Mathematics, Vol. 1627, Springer, Berlin Heidelberg New York, 197–285. Kurtz, T.G. and Xiong, J. (1999), Particle Representations for a Class of Nonlinear SPDEs. Stochast. Process Appl. 83, 103–126.
454
References
Kurtz, T.G. and Xiong, J. (2001), Numerical Solutions for a Class of SPDEs with Application to Filtering. Stochastics in Finite/Infinite Dimensions (in honor of Gopinath Kallianpur) , 233–258. Birkh¨auser, Boston. Kurtz, T.G. and Xiong, J. (2004), A stochastic evolution equation arising from the fluctuation of a class of interacting particle systems. Comm. Math. Sci. 2, 325–358. Kushner, H.J. (1967), Dynamical Equations for Optimal Nonlinear Filtering. J. Differ. Eqns., 3(2), 179–190. Ladyˇzenskaja, O.A., Solonnikov, V.A. and Ural’ceva, N.N. (1967)), Linear adn Quasilinear Equations of Parabolic Type. Nauka, Moscow, 1967 (in Russian)(English Translation (1968), American Mathematical Society). Lifshits, E.M. and Pitayevskii, L.P. (1979), Physical Kinetics. Theoretical Physics X. Nauka, Moscow (in Russian). Lions, J.L. and Magenes, E. (1972), NonHomogeneous Boundary Value Problems and Applications I. Springer, Berlin Heidelberg New York. Lions, P.L. and Souganidis, P. F. (2000), Uniqueness of Weak Solutions of Fully Nonlinear Stochastic Partial Differential Equations. C.R. Acad. Sci. Paris. t. 331, S`erie I, 783–790. Liptser, P.Sh. and Shiryayev, A.N. (1986), Theory of Martingales. Nauka, Moscow (in Russian). Liptser, P.Sh. and Shiryayev, A.N. (1974), Statistics of Random Processes. Nauka, Moscow (in Russian). Lo`eve, M. (1978), Probability Theory II. Springer, Berlin Heidelberg New York. Ma, Z.M. and R¨ockner, M. (1992), Introduction to the Theory of (NonSymmetric) Dirichlet Forms. Springer, Berlin Heidelberg New York. Marchioro, C. and Pulvirenti, M. (1982), Hydrodynamics and Vortex Theory.. Comm. Math. Phys. 84, 483–503. Marcus, M. and Mizel, V. J. (1991), Stochastic Hyperbolic Systems and the Wave Equations. Stochast. Stochast. Rep., 36, 225–244. Marcus, R. (1974), Parabolic Itˆo Equations. Trans. Am. Math. Soc. 198, 177–190. MartinL¨of, A. (1976), Limit Theorems for the Motion of a Poisson System of Independent Markovian Particles with High Density. Z. Wahrsch.verw. Gebiete 34, 205–223. Metivier, M. (1988), Stochastic Partial Differential Equations in Infinite Dimensional Spaces. Scuola Normale Superiore, Pisa. Metivier, M. and Pellaumail, J. (1980), Stochastic Integration. Adademic Press, New York. Mikulevicius, R. and Rozovsky, B.L. (1999), Martingale Problems for Stochastic PDE’s. In: Stochastic Partial Differential Equations: Six Perspectives, Carmona, R. A. and Rozovskii, B. (eds.), Mathematical Surveys and Monographs, Vol. 64, American Mathematical Society, pp. 243–325. Mikulevicius, R. and Rozovsky, B.L. (2004), Stochastic NavierStokes Equations for Turbulent Flows. Siam J. Math. Anal. 35(5), 1250–1310. Mikulevicius, R. and Rozovsky, B.L., Global L 2 Solutions of Stochastic NavierStokes Equations. Ann. Probab. 33, No 1, 137–176.
References
455
Monin, A.S. and Yaglom, A.M. (1965), Statistical Fluid Mechanics: Mechanics of Turbulence. Nauka, Moscow (English Translation, Second Printing: (1973), The MIT Press, Vols. 1 and 2.) Mueller, C. (2000), The Critical Parameter for the Heat Equation with a Noise Term to Blow up in Finite Time. Ann. Probab. 28(4), 1735–1746. Mueller, C. and Perkins, E.A. (1992) The Compact Support Property for Solutions to the Heat Equation with Noise. Probab. Th. Rel. Fields 93, 325–358. Mueller, C. and Sowers, R. (1995) Travelling Waves for the KPP Equation with Noise. J. Funct. Anal., 128, 439–498. Mueller, C. and Tribe, R. (1995) Stochastic P.D.E.’s Arising from the Long Range Contact and Long Range Voter Models. Probab. Th. Rel. Fields 102(4), 519–546. Mueller, C. and Tribe, R. (2002) Hitting Properties of a Random String. EJP 7, 1–29. Natanson, I.P. (1974), Theory of Functions of a Real Variable. Nauka, Moscow (in Russian). Neate, A.D. and Truman, A. (2006), A OneDimensional Analysis of Turbulence and its Intermittence for the dDimensional Stochastic Burgers Equations. Preprint. Nelson, E. (1972), Dynamical Theories of Brownian Motions. Princeton University Press, Princeton, N.J. Nualart, D. and Zakai, M. (1989), Generalized Brownian functionals and the solution to a stochastic partial differential equation. J. Funct. Anal. 84(2), 279–296. Oelschl¨ager, K. (1984), A Martingale Approach to the Law of Large Numbers for Weakly Interacting Stochastic Processes. Ann. Probab. 12, 458–479. Oelschl¨ager, K. (1985), A Law of Large Numbers for Moderately Interacting Diffusions. Z. Wahrscheinlichkeitstheorie verw. Gebiete 69, 279–322. Pardoux, E. (1975), Equations aux d´eriv´ees partielles stochastique non lin´eaires monotones. Etude de solutions fortes de type Itˆo. These. Pardoux, E. (1979), Stochastic Partial Differential Equations and Filtering of Diffusion Processes. Stochastics 3, 127–167. Pazy, A. (1983), Semigroups of linear operators and applications to partial differential equations (Applied Mathematical Sciences) 44, Springer, Berlin Heidelberg New York. Perkins, E. (1992), MeasureValued Branching Measure Diffusions with Spatial Interactions. Probab. Th. Rel. Fields 94, 189–245. Perkins, E. (1995), On the Martingale Problem for Interactive MeasureValued Branching Diffusions. Mem. AMS 115, No. 549, 1–89. Peszat, S. and Zabczyk, J. (2000), NonLinear Stochastic Wave and Heat Equations. Probab. Th. Rel. Fields 116, 421–443. Peszat, S. and Zabczyk, J. (2006), Stochastic Partial Differential Equations Driven by L´evy Processes. Book. Preprint. Prohorov, Yu.V. (1956), Convergence of Random Measures and Limit Theorems in Probability Theory. Th. Probab. Appl. 1(2), 157–214. Protter, P.E. (2004), Stochastic Integration and Differential Equations. Applications of Mathematics, Springer, Berlin Heidelberg New York.
456
References
Reimers, M. (1989), One Dimensional Stochastic Partial Differential Equations and the Branching Measure Diffusion. Probab. Th. Rel. Fields 81, 319–340. Roukes, M. (2001), Plenty of Room, Indeed. Scientific American, Special Issue on Nanotechnology, September 2001. Rozovsky, B.L. (1983), Stochastic Evolution Systems. Nauka, Moscow (in Russian – English Translation (1990), Kluwer Academic, Dordrecht). Schauml¨offel, K.U. (1986), Verallgemeinerte Zuf¨allige Felder und Lineare Stochastische Partielle Differentialgleichungen. Master Thesis, University of Bremen. Schmalfuss, B. (1991), LongTime Behavior of the Stochastic NavierStokes Equations. Math. Nachr. 152, 7–20. Schmuland, B. (1987), Dirichlet Forms and Infinite Dimensional OrnsteinUhlenbeck Processes. Ph.D. Thesis, Carleton University, Ottawa, Canada. Schwarz, L. (1954), Sur l’impossibilit´e de la multiplication de distributions. Comptes Rendus Hebdomadaires de S´eances de l’Academie de Sciences (Paris), 239, 847–848. Shiga, T. (1994), Two Contrasting Properties of Solutions for OneDimensional Stochastic Partial Differential Equations. Can. J. Math. 46(2), 415–437. Sinai, Ya.G. and Soloveichik, M.R. (1986), OneDimensional Classical Massive Particle in the Ideal Gas. Commun. Math. Phys. 104, 423–443. Skorokhod, A.V. (1956), Limit Theorems for Stochastic Processes. Theor. Probab. Appl.1 (269–290). Skorokhod, A.V. (1987), Asymptotic Methods of the Theory of Stochastic Differential Equations. Kiev, Naukova Dumka (in Russian). Skorokhod, A.V. (1996), On the Regularity of ManyParticle Dynamical Systems Perturbed by White Noise. Preprint. Souganidis, P.F. and Yip, N.K. (2004), Uniqueness of Motion by Mean Curvature Perturbed by Stochastic Noise. Ann. I. H. Poincar´e  AN 21, 1–23. Spohn, H. (1991), Large Scale Dynamics of Interacting Particles. Springer, Berlin Heidelberg New York. Sritharan, S. and Sundar, P., Large Deviations for the TwoDimensional Navier Stokes Equations with Multiplicative Noise. Stoch. Proc. Appl. 116, 1636–1659. Stratonovich, R.L. (1964), A New Representation of Stochastic Integrals and Equations. Vestnik MGU, Ser. 1, 1, 3–12 (in Russian). Stroock, D.W. and Varadhan, S.R.S. (1979), Multidimensional Diffusion Processes. Springer, Berlin Heidelberg New York. Suetin, P.K. (1979), Classical Orthogonal Polynomials., 2nd edition, Nauka, Moscow (in Russian). Sz´asz, D. and T´oth, B. (1986a), Bounds for the Limiting Variance of the “Heavy Particle” in R. Commun. Math. Phys. 104, 445–455. Sz´asz, D. and T´oth, B. (1986b), Towards a Unified Dynamical Theory of the Brownian Particle in an Ideal Gas. Commun. Math. Phys. 111, 41–62. Tanabe, H. (1979), Equations of Evolution. Pitman, London. Treves, F. (1967), Topological Vector Spaces, Distributions and Kernels. Academic Press, New York.
References
457
Triebel, H. (1978), Interpolation theory, function spaces, differential operators. VEB Deutscher Verlag der Wissenschaften. Berlin. Truesdell, C. and Noll, W. (1965), Encyclopedia of Physics, Vol. III/3  The Nonlinear Field Theories of Mechanics. Springer, Berlin Heidelberg New York. Tubaro, L. (1984), An Estimate of Burkholder Type for Stochastic Processes Defined by a Stochastic Integral. Stoch. Anal. Appl. 187–192. Tulpar, A., Van Tassel, P.R. and Walz, J.Y. (2006), Structuring of MacroIons Confined between LikeCharged Surfaces. Langmuir 22(6): 2876–2883. Uhlenbeck, G.E. and Ornstein, L.S. (1930), On the Theory of the Brownian Motion. Phys. Rev., 36, 823–841. ¨ unel, A.S. (1985), On the Hypoellipticity of Stochastic Partial Differential EquaUst¨ tions. Proceedings of the IFIPWG 7/1 Working Conference, LN in Control and Information Sciences, Vol. 69, Springer, Berlin Heidelberg New York. ¨ unel, A.S. (1995), An Introduction to Analysis on Wiener Space. LN in MatheUst¨ matics 1610, Springer, Berlin Heidelberg New York. Van Kampen, N.G. (1983), Stochastic Processes in Physics and Chemistry. North Holland, Amsterdam, New York. Vaillancourt, J. (1988), On the Existence of Random McKeanVlasov limits for Triangular Arrays of Exchangeable Diffusions. Stoch. Anal. Appl. 6(4), 431–446. Viot, M. (1975), Solutions faibles d’´equations aux d´eriv´ees partielles stochastique non lin´eaires. These. Vishik, M.J. and Fursikov A.V. (1980), Mathematical Problems of Statistical Hydromechanics. Nauka, Moscow (Tranlation into Enlish (1988), Kluwer, Dordrecht.,) Walsh, J.B. (1981), A Stochastic Model of Neural Response. Adv. Appl. Prob. 13, 231–281. Walsh, J.B. (1986), An Introduction to Stochastic Partial Differential Equations. Ecole d’Et´e de Probabilit´e de Saint Fleur XIV. Lecture Notes in Math. 1180. Springer, Berlin Heidelberg New York, 265–439. Wang, H.(1995), Interacting Branching Particles System and Superprocesses. Ph.D. Thesis. Carleton University, Ottawa. Wang, W., Cao, D. and Duan J. (2005), Effective Macroscopic Dynamics of Stochastic Partial Differential Equations in Perforated Domains. SIAM. J. Math. Ana Vol 38, No 5, 1508–1527. Watanbe, S. (1984), Lectures on Stochastic Differential Equations and Malliavin Calculus. Tata Institute of Fundamental Research, Bombay. Weinan, E. Mattingly, J. and Sinai, Ya. (2001), Gibbsian Dynamics and Ergodicity for the Stochastically Forced NavierStokes Equation. Comm. Math. Phys. 224, 83–106. Wong, E. and Zakai, M. (1965), On the Relation between Ordinary and Stochastic Differential Equations, Int. J. Eng. Sci. 3, 213–229. Yaglom, A.M. (1957), Some Classes of Random Fields in nDimensional Space, Related to Stationary Random Processes, Theor. Probab. Appl. II(3), 273–320. Yip, N.K. (1998), Stochastic Motion by Mean Curvature. Arch. Rational Mech. Anal. 144, 313–355.
458
References
Yosida, K. (1968), Functional Analysis, Second Edition. Springer, Berlin Heidelberg New York. Zabczyk, J. (1977), Linear Stochastic Systems in Hilbert Spaces: Spectral Properties and Limit Behavior. Institute of Mathematics, Polish Academy of Sciences, Report No. 236. Zakai, M. (1969), On the Optimal Filtering of Diffusion Processes. Z. Wahrschein. Verw. Geb.11, 230–243.
Stochastic Modelling and Applied Probability formerly: Applications of Mathematics
(continued from page ii)
Asmussen, Applied Probability and Queues (2nd. ed. 2003) Robert, Stochastic Networks and Queues (2003) Glasserman, Monte Carlo Methods in Financial Engineering (2004) Sethi/Zhang/Zhang, AverageCost Control of Stochastic Manufacturing Systems (2005) Yin/Zhang, DiscreteTime Markov Chains (2005) Fouque/Garnier/Papanicolaou/Sølna, Wave Propagation and Time Reversal in Random Layered Media (2007) 57 Asmussen/Glynn, Stochastic Simulation: Algorithms and Analysis (2007) 58 K Kotelenez, Stochastic Ordinary and Stochastic Partial Differential Equations: Transition from Microscopic to Macroscopic Equations (2008) 51 52 53 54 55 56