Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Applications
Kazufumi Ito∗
October 15, 2013
Abstract
In recent years increasing interests and considerable researches have been given to
the fractional differential equations both in time and space variables. These are due to
the applications of the fractional differential operators to problems in a wide areas of
physics and engineering science and a rapid development of the corresponding theory.
Motivating examples include the so-called continuous time random walk process and the
Levy process model for the mathematical finance. Basset integral is appearing in the
equation of motion of a particle moving through a fluid. A fractional diffusion equation
is derived as a homogenization of heterogeneous groundwater flow. In this lecture we
develop solution methods based on the linear and nonlinear semigroup theory and
apply it to solve the corresponding inverse and optimal control problems. The theory
is applied to concrete examples including fractional diffusion equation, Navier-Stokes
equations and conservation laws. For the linear case we develop the operator theoretic
representation of solutions and the sectorial property of the fractional operator in time
is used to establish the regularity and asymptotic of the solutions. The property and
stability of the solutions as well as numerical integration methods are discussed. The
lecture also covers the basic theory and application of the so-called Crandall-Ligget
theory and the DS-approximation theory developed by Kobayashi-Kobayashi-Oharu
for evolution operator and the semi-linear theory based on the sectorial estimates of
the fractional equation.
1 Introduction
In this monograph we consider the fractional power equation of the form
(t − s)−α 0
Z t
α
Dt u = u (s) ds = Au(t) + f (t), u(0) = x (1.1) fra
0 Γ(1 − α)
1
will focus on fractional power equations in time but A may represent the fractional
power operator in space such as fractional power Laplacean −(−∆)β (also, see Section
). In general we consider equations of the form
Z t
g(t − s)u0 (s) ds = Au(t) + f (t), u(0) = x, (1.2) frag
0
where we assume g(s) ≥ fra 0 is monotonically deceasing and integrable on any finite
frag interval (0, R). Equation (1.1) is the special case of with
t−α
g(t) = g1−α (t) = . (1.3) gal
Γ(1 − α)
As shown in Appendix, such an equation arises in the continuous time random walk
model for MK
groundwater movement in naturally fractured SGM and heterogeneous porous
aquifers [3] and tick-by-tick dynamics of financial markets [5]. Boussinesq (1885) and
Basset (1888) found that the force F on an accelerating spherical particle in a viscous
fluid is given by
− v 0 (s)
Z t Du
3 √
F = D2 πρc µc ( Ds√ ds,
2 0 t−s
D
where D is the particle diameter, Ds is the material derivative, and u and v are the
fluid and particle velocity vectors, respectively. Thus, one can the Basset equation as
frag t−1/2
(1): with g = δ0 + k Γ(1/2) . i.e.
1
u0 (t) + kDt2 u = Au(t) + f (t), u(0) = x.
Or, in general
1
t−α
Z
g(t) = e−βt dµ(α)
0 Γ(1 − α)
where µ is a positive measure on (0, 1]. The exponential decay model is
t−α
g(t) = e−βt
Γ(1 − α)
we have
Z t
d t d t
Z Z
0
g(t − s)u (s) dx = g(t − s)u(s) ds − g(t)u(0). = g(t − s)u(s) ds, (1.4) trans
0 dt 0 dt −∞
2
where we set u(s) = x, s ≤ 0. Equivalently,
Z t Z t
0 d
g(t − s)u (s) dx = g(t − s)(u(s) − u(0)) ds.
0 dt 0
frag
Thus, (1) is written as the fractional differential equation of the Riemann-Liouville
form:
d t (t − s)−α (t − s)−α
Z
u(s) ds = Au(t) + f (t) + u(0),
dt 0 Γ(1 − α) Γ(1 − α)
or
d t
Z
g(t − s)(u(s) − u(0)) ds = Au(t) + f (t).
dt 0
fra
There are several approaches to define the solution to (1.1). One uses the Mitag-
Leffler function Eα,β (t) defined by
∞
X λn tnα
Eα,β (λt) = . (1.5) ML
Γ(nα + β)
n=0
The other approach is the integral equation approach. Let Jtα be the operator
defined by Z t
α (t − s)α−1
Jt φ = φ(s)) ds.
0 Γ(α)
Then, Jtα Dtα u = u(t) − x, ; t ≥ 0 and we obtain the integral equation for u;
t
(t − s)α−1
Z
u(t) = x + (Au(s) + f (s)) ds,
0 Γ(α)
3
fde fra
1
g(θ) = Γ(1−α) |θ|−α , then (1.6) reduces to (1.1). That is, if θ → u(t + θ) is absolute
tars
continuous, it follows from (??) (see, Section 2.2 for the precise discussions) that
Z 0 Z t
0
g(θ)u (t + θ) dθ = g(t − s)u0 (s) ds.
−∞ 0
We then embed the solution u(t) = z(t, 0) in the state (history) space
in Z with domain
Z 0
0
dom (A(t)) = {φ ∈ Z : φ ∈ Z and g(θ)φ0 (θ) dθ = Aφ(0)+f (t), φ(0) ∈ dom(A)}.
−∞
frag evol
(1.9) evol1
Dynamics (1) is embedded in (1.7) as the non-local boundary value condition as θ =
0+ for the first order differential frag
operator A(t). We analyze the well-posedness
evol
and
the property of of solutions to (1) based on the semigroup generated by (1.7), i.e.,
show that the solution map (x, f ) ∈ X × C(0, T ; X) → u(t) ∈ C(0, T ; X) exists and
continuous. It will be shown that if A is dissipative and maximal monotone in X,
then A(t) is dissipative and maximal monotone in Z. We then evol
use the semigroup
generation
frag
theory to define the solution z(t) ∈ C(0, T ; Z) to (1.7) and the solution to
(1) by u(t) = z(t, 0) ∈ C(0, T ; X). In this
farg
way we can definefrathe solution to a more
general class of equations of the form (??). In the case of (1.1) with a closed linear
operator A, we have
Z t Z t
(t − s)α−1 (t − s)α−1
u(t) = x + A u(s) ds + f (s) ds.
0 Γ(α) 0 Γ(α)
for all x ∈ X and f ∈ C(0, T ; X), i.e.,
Z t
(t − s)α−1
u(s) ds ∈ C(0, T ; dom (A)).
0 Γ(α)
For the case
fra
of a closed
frag
linear operator A, we also develop the
fra
operator theoretic
approach to (1.1) and (1). If we take the Laplace transform of (1.1)
λα û = λα−1 x + fˆ
and thus
û = (λα I − A)−1 (λα−1 x + fˆ)
assuming A is maximal monotone. Here, for 0 < α < 1 there exists θ > 0 such that
Mα π
|(λα I − A)−1 | ≤ on the sector Σθ = {z ∈ C : arg(z) ≤ + θ} ∩ {z 6= 0}.
|λ|α 2
4
Thus, we have the representation of the solution operator u(t) = S(t)x:
Z
1
S(t)x = eλt (λα I − A)−1 λα−1 x dλ.
2πi Γ
where Γγ,δ be the integration path defined by
π π
Γ± = {z ∈ C : |z| ≥ δ, arg(z) = ±( +γ)}, Γ0 = {z ∈ C : |z| = δ, |arg(z)| ≤ +γ}.
2 2
For t > 0 define P (t) ∈ L(X) by
Z
1
P (t)x = eλt (λα I − A)−1 x dλ.
2πi Γ
We will analyze the solution properties based on this operator representation of S(t)
and P (t) and establish the regularity and asymptotic property as t → ∞ of S(t) and
P (t). For the sectorial operator A we analyze the properties of S(t) and P (t) based on
the fractional operator calculus.
For the semilinear equation with Au = A0 u + F (u), we define the mild solution by
the solution representation;
Z t
u(t) = S(t)x + P (t − s)(F (u(s) + f (s)) ds. (1.11) semil
0
5
[10] Local solutions to nonlinear fractional power ODEs, Section 10.
[11] Local solutions to semi-linear fractional power PDEs, Section 11.
[12] Examples, fractional diffusion, conservation law, Hamilto-Jacobi and Navier-
Stokes equations, Section 12.
[13] Nonlocal and fractional PDEs in space, Section 13
[14] Eigenvalue Problems for fractional operator, Section 14
[15] CTRM model and Fractional diffusions via homogenization, Appendix.
Aφ = φ0 (θ)
with domain
Z 0
0
dom (A) = {φ ∈ Z : φ ∈ Z and g(θ)φ0 (θ) dθ = Aφ(0), φ(0) ∈ dom(A)}.
−∞
frag
The evolution equation (1) for u is embedded into the non-local boundary condition
at θ = 0+ in the domain of the closed operator A.
Proof: Define Z θ
1
g (θ) = g(θ) dθ
θ−
For φ ∈ dom(A)
Z 0 Z 0
0
(Aφ, φ)Z = g(θ)(φ (θ), φ(θ) − φ(0))H dθ + ( g(θ)φ0 (θ) dθ, φ(0))H ,
−∞ −∞
6
where
Z 0 Z 0
0
g(θ)(φ (θ), φ(θ) − φ(0))H = lim g (θ)(φ0 (θ), φ(θ) − φ(0))H dθ
−∞ →0+ −∞
(2.1) sta0
Z 0
1
=− lim g0 (θ)|φ(θ) − φ(0)|2 dθ ≤ 0,
2 →0+ −∞
then
λψ − ψ 0 = f ∈ Z, ψ(0) = 0,
1
and |ψ|Z ≤ |f |Z . Let
λ
Z −δ Z δ
0
I= g(θ)ψ (θ) dθ = g(−δ)ψ(−δ) − g(−R)ψ(−R) − g 0 (θ)ψ(θ) dθ. (2.3) bdd
−R −R
Let Z 0
∆(λ) = λ eλθ g(θ) dθ. (2.4) Delta
−∞
Since A is m-dissipative, λ φ − Aφ = f has the solution
where Z 0
φ(0) = (∆(λ) I − A)−1 g 0 (θ)ψ(θ) dθ.
−∞
IK
The theorem now follows from the Lumer-Phillips theorem [2].
7
2.2 State space Z = C((−∞, 0]; H)
Let X = H be a Hilbert space and consider the state space
Z = C((−∞, 0]; H).
Define a linear operator A by
Aφ = φ0
with domain
Z 0
dom(A) = {φ0 ∈ Z : φ(0) ∈ dom(A), g(θ)φ0 (θ) dθ = Aφ(0)}
−∞
Then,
0 0
g(θ) − g(θ − )
Z Z
0
( g (θ)φ (θ) dθ, φ(0)) = ( (φ(0) − φ(θ)) dθ, φ(0))
−∞ −∞
0
g(θ) − g(θ − )
Z
= ((φ(0), φ(0)) − (φ(θ), φ(0))) dθ (2.5) sta
−∞
0
g(θ) − g(θ − )
Z
≥ (|φ(0)|2 − |φ(θ)||φ(0)|) dθ
−∞
Note that Z 0
lim ( g (θ)φ0 (θ) dθ, φ(0)) = (Aφ(0), φ(0)) (2.6) sta1
→0+ −∞
sta sta1
Suppose |φ(0)| > |φ(θ)|, then it follows from (2.5)–(2.6) that
(Aφ(0), φ(0)) > 0
which is a contradiction to the fact that A is dissipative . That is, max |φ| = |φ(θ0 )|
for some θ0 < 0 and
|λφ − Aφ|Z ≥ |λφ(θ0 ) − φ0 (θ0 )|Z ≥ λ |φ(θ0 )| = λ |φ|Z
bdd
since (φ0 (θ0 ), φ(θ0 ))H = 0. Thus, A is dissipative. Moreover, for λ > 0 From (2.3)
|I| ≤ 2(g(−δ) + g(−R))|ψ|Z
and thus φ = eλθ φ(0) + ψ ∈ dom (A) satisfies
(λ I − A)φ = f in Z,
where Z 0 Z 0
φ(0) = (∆(λ) − A)−1 g(θ)(f (θ) − λ eλ(θ−ξ) f (ξ) dξ) dθ
−∞ θ
(2.7) resl
Z 0
φ = (λ I − Aφ )−1 f = eλθ φ(0) + eλ(θ−ξ) f (ξ) dξ.
θ
Hence, we have
thm2.2 Theorem 2.2. Assume A is m-dissipative. Then, A is dissipative and R(λ I −A) = Z
for λ > 0. Thus, A generates the C0 -semigroup T (t) on Z = C((−∞, 0]; H).
8
2.3 Banach space-valued solution
In this case section we discuss the case when X is a Banach space and A is m-dissipative
linear operator in X. Let X ∗ is the dual space of X, h·, ·iX×X ∗ denote the dual product
and F be the duality mapping
Then , A is dissipative if and only if for all x ∈ dom (A) there exists a x∗ ∈ F (x), the
such that
hAx, x∗ i ≤ 0
and we assume
range(λ I − A) = X for all λ > 0.
Then we have the generation theory:
Proof: First we show that A is dissipative. For φ ∈ dom (A) suppose |φ(0)| > |φ(θ)|
for all θ < 0. For all x∗ ∈ F (φ(0))
Z 0
h g (θ)(φ0 )dθ, x∗ i
−∞
0
g(θ) − g(θ − )
Z
=h hφ(θ) − φ(0)), x∗ i dθ ≤ 0
−∞
since
hφ(θ) − φ(0), x∗ i ≤ (|φ(θ)| − |φ(0)|)|φ(0)| < 0, θ < 0.
Thus, Z 0
h g(θ)φ0 dθ, x∗ i < 0. (2.9) ine0
−∞
hAx, x∗ i ≥ 0
ine0
which contradicts to (2.9). Thus, there exists θ0 such that |φ(θ0 )| = |φ|Z . Since
hφ(θ), x∗ i ≤ |φ(θ)| for x∗ ∈ F (φ(θ0 )), θ → hφ(θ), x∗ i attains the maximum at θ0 and
thus hφ0 (θ0 ), x∗ i = 0 Hence,
9
frag
2.4 Mild solution to (1)
resl
Consider the case with the initial value φ(θ) = x ∈ Z. It follows from (2.7) that for
µ>0
θ
(I − µ A)−1 x = φµ , φµ (θ) = x + e µ (φµ (0) − x)
with
1 1
φµ (0) = (∆( ) I − A)−1 ∆( )x
µ µ
since Z 0
λ eλ(θ−ξ) dξ − 1 = eλθ .
θ
Note that
1 −1 1
φµ (0) − x = (I − 1 A) Ax
∆( µ ) ∆( µ1 )
Thus,
1
|φµ (0) − x| ≤ |Ax|X → 0 as µ → 0
∆( µ1 )
for all x ∈ dom(A). Since φµ ∈ dom(A), zµ = uµ (t + ·) = T (t)φµ is a strong solution
to
d
zµ (t) = Azµ (t), zµ (0) = φµ .
dt
That is, uµ ∈ C 1 ((−∞, T ]; X) satisfies
Z t Z 0
0
g(t − s)uµ (s) ds = Auµ (t) − g(−t + θ)φ0µ (θ) dθ.
0 −∞
where Z 0
| g(−t + θ)φ0µ (θ) dθ| ≤ g(t)|φµ (0) − x| → 0 as µ → 0+ .
−∞
fde
Since Dtα is closed on C(0, T ; X), u(t) = limµ→0+ uµ (t) satisfies (1.6);
Dtα u = Au(t)
Thus,
Z τ Z t Z τ Z τ Z τ
g ∗ (τ −s) g(t−s)u0 (s) ds = u0 (s) g ∗ (τ −t)g(t−s) dtds = u0 ds = u(τ )−u(0).
0 0 0 s 0
(2.12) gs0
Since AT (t) = T (t)A, Auµ (t) → Au(t) in C(0, T ; X) and the limit u(t) = limµ→0+ uµ (t)
satisfies Z t
u(t) = x + A g ∗ (t − s)u(s) ds. (2.13) int1
0
frag
Since dom(A) is dense in X and u(t) ∈ C(0, T ; X) defines the mild solution to (1) with
f = 0.
10
fra tα−1
In the case of (1.1), g ∗ (t) = , i.e,
Γ(α)
τ
(τ − t)α−1 (t − s)−α
Z
=1 (2.14) J
0 Γ(α) Γ(1 − α)
Proof: The first assertion follows since dom (A) is dense in X and A is a closed operator
in X. The second one follows since
d t (t − s)α−1
Z t
(t − s)α−1 0 tα−1
Z
0
u (t) = Au(s) ds = Au (s) ds + Ax.
dt 0 Γ(α) 0 Γ(α) Γ(α)
Aφ = φ0
with domain
Z 0
0
dom(A) = {φ ∈ Z : φ(0) ∈ dom(A), g(θ)φ0 (θ) dθ ∈ Aφ(0)}.
−∞
thm3.1 Theorem 3.1. Assume A is dissipative and Range(λ I − A) for all sufficiently small
λ > 0. Then, A is dissipative and Range(λ I − A) = Z for all sufficiently small λ > 0.
Thus A generates the nonlinear semigroup of contraction on Z.
11
Proof: For φ1 , φ2 ∈ dom (A) suppose |φ1 (0) − φ2 (0)| > |φ1 (θ) − φ2 (θ)| for all θ < 0.
For all x∗ ∈ F (φ1 (0) − φ2 (0))
Z 0
h g (θ)(φ01 − φ02 )dθ, x∗ i
−∞
0
g(θ) − g(θ − )
Z
=h hφ1 (θ) − φ2 (θ) − (φ1 (0) − φ2 (0)), x∗ i dθ ≤ 0
−∞
since
hφ1 (θ)−φ2 (θ)−(φ1 (0)−φ2 (0), x∗ i ≤ (|φ1 (θ)−φ2 (θ)|−|φ1 (0)−φ2 (0))|)|φ1 (0)−φ2 (0)| < 0, θ < 0.
Thus, Z 0
h g(θ)(φ01 − φ02 )dθ, x∗ i < 0. (3.1) ine
−∞
But, since for y1 ∈ Aφ1 (0), y2 ∈ Aφ2 (0) there exists a x∗ ∈ F (φ1 (0) − φ2 (0)) such that
hy1 − y2 , x∗ i ≥ 0
ine
which contradicts to (3.1). Thus, there exists θ0 such that |φ1 (θ0 )−φ2 (θ0 )| = |φ1 −φ2 |Z
and thus hφ0 (θ0 ), x∗ i = 0 for all x∗ ∈ F (φ1 (θ0 ) − φ2 (θ0 )). Thus,
|λ (φ1 − φ2 ) − (φ0 − φ02 )|Z ≥ hλ (φ1 (θ0 ) − φ2 (θ0 )) − (φ01 (θ0 ) − φ02 (θ0 )), x∗ i
(3.2) ineq
= λ |φ1 (θ0 ) − φ2 (θ0 )| = λ |φ1 − φ2 |Z .
(I − s A)−1 C ⊂ C
12
thm3.2 Theorem 3.2. A is cone preserving and T (t)C ⊂ C.
Proof: From 1
θ
φ = (I − µ A)−1 f = e µ φ(0) + ψ(θ)
with Z 0
1 −1
φ(0) = (∆( ) − A) ( g 0 (θ)ψ(θ) dθ)
µ −∞
Z 0
1 1
(θ−ξ)
ψ(θ) = eµ f (ξ) dξ
µ θ
Thus, since g 0 ≥ 0 if f ∈ C, then φ ∈ C.
We assume a family of dissipative operators A(t) ⊂ X×X, t ∈ [0, T ] satisfy dom (A(t)) =
X. Define the operator A(t) in Z = C((−∞, 0]; X) by
A(t)φ = φ0
with domain Z 0
dom (A(t)) = { g(θ)φ0 (θ) dθ ∈ A(t)φ(0)}.
−∞
For λ > 0 define the resolvent
|Aλ (t1 )z − Aλ (t2 )z| ≤ λα−1 L |f (t1 ) − f (t2 )|, (4.3) cont
13
For λ > 0 let {zkλ } be the sequence generated by
That is, the product formula zkλ = Πm i=1 Jλ (ti )z defines an approximation sequence and
satisfies
|Πm m
i=1 Jλ (ti )z1 − Πi=1 Jλ (ti )z2 | ≤ |z1 − z2 |. (4.4) prod
cont
From (8.4)
|Aλ (tλk )zkλ | = |Aλ (tλk )Jλ (tλk )zk−1
λ | ≤ |A (tλ )z λ |
λ k k−1
we have
ak = ak−1 + λα−1 bk
and thus
|Aλ (tλk )| ≤ M0 λα−1 (4.5) bound
for all k and λ.
µ
Let λ = 2−n , µ = 2−m and N = 2m−n with tλk = k λ. For 1 ≤ j ≤ N define ẑiN +j
by
µ λ µ
ẑiN +j = Jµ (t(i+1)N )ẑiN +j−1 .
A.2
It follows from (4.2) that
µ µ µ µ α µ µ
|ziN +j − ẑiN +j | ≤ |ziN +j−1 − ẑiN +j−1 | + µ |f (t(i+1)N ) − f (tiN +j )|L.
Then,
|zµ (t) − ẑµ (t)| ≤ 2(1−γ) m−n Lλ. (4.6) err1
It follows from Theorem 5.3 (Crandall-Liggett theorem) that
14
2−α
if m ≤ n. Let us define the sequence {mk } by
1−γ
2−α
mk = [ mk−1 ], m0 = n.
1−γ
Then, by induction in k
1−α+γ
where µk = 2−mk . Since mk − mk−1 ≥ m0 > 0,
1−γ
|zµk − zµ0 | ≤ M λα
for some M > 0 and thus {zλ }, λ = 2−n is a Cauchy sequence. Consequently,
nonle
z(t) = limλ→0 exits in C(0, T ; Z) which defines the solution to (4.1) with z(0) =
φ ∈ dom(A(0)) and
|zλ (t) − z(t)| ≤ M λα .
prod
It follows from (4.4) that
thm4.1 Theorem 4.1. Assume (A2) and f is of bounded variation and Hölder continuous of
order 1 − α + γ, γ > 0 on [0, T ]. Then,
[t/λ]
z(t) = lim Πi=1 Jλ (tλi )z(0)
λ→0
15
For F (t)x = A(0)A(t)x − A(t)A(0)x we have
Thus, if A(0)x(0) ∈ X, one can show that there exists M > 0 such that
Assume that A is a closed densely defined linear operator A in the Banach space X
and there exist M ≥ 1 and ω0 ∈ R such that
M
|(λ I − A)−1 | ≤ for all Re λω0 . (5.2) res
Re λ − ω0
res
For example, (5.2) holds if A generates a C0 semigroup of G(M, ω0 ) type on X. Define
the Yoshida approximation Aµ ∈ L(X) of A by
and thus
ûµ = (∆(λ) I − Aµ )−1 λ−1 ∆(λ)x + (∆(λ) I − Aµ )−1 fˆ.
ω0
Let f = 0. Since Re σ(Aµ ) ≤ < γ, we have
1 − µω0
Z γ+i∞
1
uµ (t) = Sµ (t)x = eλt (∆(λ) − Aµ )−1 λ−1 ∆(λ)x dλ.
2πi γ−i∞
Note that
∆(λ)(∆(λ) I − A)−1 = I + (∆(λ) I − A)−1 A (5.4) id
16
and in general
n−1
X
∆(λ)(∆(λ) I − A)−1 = ∆(λ)−k Ak + (∆(λ) I − A)−1 ∆(λ)−n An . (5.5) ser
k=0
Since
γ+i∞
eλt
Z
1
dλ = 1
2πi γ−i∞ λ
and Z γ+i∞
|λ−1 ∆(λ)−1 | dλ < ∞,
γ−i∞
we have
|Sµ (t)x| ≤ M |Ax|,
id
uniformly in µ > 0. Since from (5.4)
1
∆(λ)λ−1 (∆(λ) I − A)−1 = + λ−1 A(∆(λ) I − A)−1 ,
λ
if L(∆(λ)−1 ) = g ∗ we have
Z t
uµ (t) = x + g ∗ (t − s) Aµ Sµ (s)x ds.
0
Moreover, since
µ
(∆(λ) I − Aµ )−1 x − (∆(λ) I − A)−1 x = (ν I − A)−1 (∆(λ) I − A)−1 A2 x,
1 + µ∆(λ)
∆(λ)
where ν = , {uµ (t)} is Cauchy in C(0, T ; X) provided that x ∈ dom(A3 ).
1 + µ∆(λ)
Letting µ → 0+ , we obtain
thm5.1 Theorem 5.1. For x ∈ dom(A2 )
Z γ+i∞
1
u(t) = S(t)x = eλt (∆(λ) I − A)−1 λ−1 ∆(λ)x dλ
2πi γ−i∞
(5.6) res1
Z t
u(t) = x + g ∗ (t − s)AS(s)x ds.
0
17
and
t
(t − s)α−1
Z
u(t) = x + AS(t)x dx.
0 Γ(α)
ser
Proof: The first assertion follows from (5.5). For the Caputo case ∆(λ) = λα and
tα−1
g ∗ (t) = .
Γ(α)
d t
Z
g(t − s)u(s) ds = Au(t) + f (t) (5.8) SL
dt 0
with Z t
( g(t − s)u(s) ds)(0+ ) = y.
0
Taking the Laplace transform, we have
∆(λ)û = y + Aû + fˆ
and thus
û = (∆(λ) I − A)−1 (y + fˆ) (5.9) Lap
Note that
∆(λ)(∆(λ) I − A)−1 = I + (∆(λ) I − A)−1 A
Let the linear operator P (t)
Z
1
P (t)y = eλt (∆(λ) I − A)−1 y dλ,
2πi Γ
Lap
for y ∈ dom(A). Hence, using the same arguments as above, from (5.9) we have;
SL
th5.2 Theorem 5.2. The solution to (5.8) is given by
Z t
u(t) = P (t)y + P (t − s)f (s) ds
0
for y ∈ dom(A) and f ∈ L2 (0, T ; dom(A), and for the case of fractional derivative
|θ|−α
g = g1−α = ;
Γ(1 − α)
n−1 γ+i∞
tkα
Z
X 1
P (t)y = t α−1
Ak y + eλt (λα I − A)−1 λ−nα An y dλ.
Γ(kα + α) 2πi γ−i∞
k=0
for y ∈ dom(An+2 ).
18
Proof: The second assertion follows from
Z γ+i∞
1 tα−1
λ−α dλ =
2πi γ−i∞ Γ(α)
and
tα−1
L−1 (λα−1 ) = .
Γ(α)
19
Then, for 0 < α < 1 there exists θ0 > 0 such that
Mα π
|(∆(λ) I − A)−1 | ≤ on Σθ = {z ∈ C : arg(z) < + θ0 } ∩ {z 6= 0}
|λ|α 2
since
Re λα ≥ |λ|α cos(αθ) for λ = |λ|eiθ .
Let Γγ,δ be the integration path defined by
π π
Γ± = {z ∈ C : |z| ≥ δ, arg(z) = ±( + θ)}, Γ0 = {z ∈ C : |z| = δ, |arg(z)| ≤ + θ}
2 2
for some δ > 0 and 0 < θ ≤ θ0 . Then, the solution map S(t) : x ∈ X → u(t) ∈ X is
given by Z
1
S(t)x = eλt (∆(λ) I − A)−1 ∆(λ)λ−1 x dλ, (5.13) Rep1
2πi Γ
res1
using the Cauchy integral representation (5.6) and the analytic continuation. On Γ±
M −r sin θ t
|eλt (∆(λ) I − A)−1 ∆(λ)λ−1 | ≤ e
r
for λ = r(cos(θ) + i sin(θ)), r ≥ δ. On Γ0
M δ sin φ t
|eiλt (∆(λ) I − A)−1 ∆(λ)λ−1 | ≤ e
δ
π Rep1
for λ = δ(cos(φ) + i sin(φ)), |φ| ≤ 2 + θ. Hence (5.13) holds for all x ∈ X. Let
P (t) ∈ L(X), t > 0 be
Z
1
P (t)x = eλt (∆(λ) I − A)−1 x dλ.
2πi Γ
Since
∆(λ) (∆(λ) I − A)−1 = I + A(∆(λ) I − A)−1 ,
if x ∈ dom(A), then u(t) = S(t)x ∈ C(0, T ; X) for x ∈ dom(A) satisfies
Z
d 1
S(t)x = eλt (∆(λ) I − A)−1 ∆(λ)x dλ = P (t)Ax.
dt 2πi Γ
20
thm5.3 Theorem 5.3. Assume 0 ∈ ρ(A). Then, for some C > 0
d C
| S(t)x| ≤ |x|,
dt t
C
|P (t)x| ≤ |x|,
t1−α
Z
1 1 1
AS(t)x = eλt ((∆(λ) I − A)−1 ∆(λ)2 − ∆(λ) )x dλ.
2πi Γ λ λ
Proof: Since 0 ∈ ρ(A), one can let δ = 0 for the integral path Γ and
|∆(λ)(∆(λ) I − A)−1 | ≤ M on Γ.
Thus,
Z Z ∞
1 M M
| eλt (∆(λ) I − A)−1 ∆(λ)x dλ| ≤ e−r sin θ t dr ≤ .
2πi Γ π 0 π sin θ t
Z Z ∞
1 M M
|P (t)| = | eλt (∆(λ) I−A)−1 dλ| ≤ r−α e−r sin θ t dr ≤ Γ(1−α)(sin θ t)α−1 .
2πi Γ π 0 π
thm5.4 Theorem 5.4. For t > 0 R(S(t)) ⊂ R(A−1 ). If R(A−1 ) is pre-compact and S(t) is
injective, then R(S(t)) = R(A−1 ).
Proof: Since
A(λα I − A)−1 = λα (λα I − A)−1 − I, (5.14) Eq
we have
Z
1 1
AS(t)x = eλt λ2α−1 (λα I − A)−1 x dλ − t−α x = Kx − g1−α (t)x
2πi Γ Γ(1 − α)
where Z
1
Kx = eλt λ2α−1 (λα I − A)−1 x dλ
2πi Γ
t−2α
Z
1
AK = eλt λ3α−1 (λα I − A)−1 x dλ − ∈ L(X),
2πi Γ Γ(1 − 2α)
21
cor5.3 Corollary 5.3. For t > 0
n−1
t−kα
Z
X 1
S(t) = − A−k + eλt λnα−1 (λα I − A)−1 A−n dλ. (5.15) ser1
Γ(1 − kα) 2πi Γ
k=1
For t > 0 sufficiently large AS(t) = K − g1−α (t) I ∈ L(X) is bounded invertible, i.e,
there exits a constant c such that
|x| ≤ c |AS(t)x|.
Moreover, for x ∈ X
t−α
|S(t)x| ∼ |A−1 x|.
Γ(1 − α)
sere
Proof: Equation (??) follows from
n−1
X
An (λα I − A)−1 = λnα (λα I − A)−1 − λkα Ak .
k=0
Thus, g1−α (t)−|K| > 0 for sufficiently large t > 0. and |(g1−α (t) I −K)−1 | ≤ (g1−α (t)−
|K|)−1 .
Similarly, we have
M −(1−α)
|P (t)| ≤ t .
Γ(α)
Moreover, assume A is a sectorial operator, i.e., there exist M > 0, θ0 > 0 such that
M π
|(z I − A)−1 | ≤ on Σθ0 = {z ∈ C : arg(z) ≤ + θ0 } ∩ {z 6= 0}
|z| 2
Then, Z
β
A x= λβ (z I − A)−1 x
Γ
and for 0 ≤ β ≤ 1
M
|Aβ P (t)| ≤ . (5.16) sec-frac
t1−α+βα
Eq
Proof: It follows from (5.14). The last assertion uses |Aβ x| ≤ M |Ax|β |x|1−β for
x ∈ dom(A).
22
5.4 Basset equation
Consider the Basset equation
In this case
∆(λ) = λ + k λα .
Rep1
Assume A a sectorial operator so that (5.13) holds. Since
we have
AS(t)x = Kx − k g1−α (t)x,
where Z
1
Kx = eλt λ−1 ∆(λ)2 (∆(λ) I − A)−1 x dλ.
2πi Γ
Moreover, we have the estimate:
6 Series Solution
Define the operator:
t
(t − s)α−1
Z
Jtα φ = φ(s) ds
0 Γ(α)
Since Jt1−α Jtα = Jt1 , we have
Thus,
t
(t − s)α−1
Z
S(t)x = x + A S(s)x ds. (6.2) equi1
0 Γ(α)
equi
Since Au ∈ C(0, T ; H) for x ∈ dom(A) it follows from (6.1) that
u(t) − x 1
lim α
= Ax
t→0+ t Γ(α + 1)
23
Moreover, we have
1
u(t) − x − Γ(α+1) Ax 1
lim = A2 x
t→0+ t2α Γ(2α + 1)
we look for an second order approximation of the form
with
1 1
c1 − b1 = , b21 − c1 b1 = ,
Γ(α + 1) Γ(2α + 1)
Thus, we obtain
Γ(α + 1) 1 Γ(α + 1)
b1 = − , c1 = − .
Γ(2α + 1) Γ(α + 1) Γ(2α + 1)
Xn n
X
kα k −1
u(t) ∼ ( bk t A ) (I + ak tkα Ak )x.
k=0 k=1
we have
n
X X n
X
kα jα
( bk t )( cj t ) = a` t`α
k=0 j=0 `=0
Next, we consider the series expansion in terms of the resolvent (λα I − A)−1 . The
Post-Widder inversion theory is given as:
FW Theorem 6.1. Let u(t) be a X-valued continuous function on t ≥ 0 such that u(t) =
O(eγt ) as t → ∞ for some γ and û be the Laplace transform of u(t). Then,
(−1)n n n+1 ∂ n n
u(t) = lim ( n û)( ),
n→∞∞ n! t ∂λ t
uniformly on any compact sets of t > 0.
24
thm5.6 Theorem 6.2. If A generates a C0 semigroup on X, then
n+1
∂ n α−1 α X
n
(λ (λ I − A)−1 ) = (−1)−(n+1) λ−(n + 1) bαk,n+1 (λα (λα I − A)−1 )k
∂λ
k=1
n+1
1 X α t
S(t)x = lim bk,n+1 (I − ( )α A)−k x
n→∞ n! n
k=1
n+1
∂ n α−1 α −1
X
| (λ (λ I − A) )| ≤ bαk,n+1 λkα−n−1 |(λα I − A)−k |
∂λn
k=1
n+1
X λkα−n−1 n ∂
n λα−1
≤M bαk,n+1 = M (−1) ( )
(λα − w)k ∂λn λα − ω
k=1
Since Z ∞
λα−1
= e−λt Eα,1 (t) dt,
λα − ω 0
n α−1 Z ∞
∂ λ n!
(−1)n n ( α )= tn e−λt Eα,1 (t) dt ≤ C 1
∂λ λ − ω 0 (λ − ω α )n+1
1
since Eα,1 (ωtα ) ≤ Ce α t . Thus,
∂ n α−1 α n!
| n
(λ (λ I − A)−1 )| ≤ CM 1
∂λ (λ − ω α )n+1
25
where the convergence is uniform on bounded intervals of t > 0. The solution to (??)
is given by Z t
u(t) = S(t)x + P (t − s)f (s) ds.
0
corr5.6 Corollary 6.2. For For the Caputo equation with A ∈ L(X)
∞
α
X An tnα
S(t) = Eα,1 (At ) =
Γ(nα + 1)
n=0
and
∞
X An tnα
P (t) = tα−1 Eα,α (Atα ) = tα−1 .
Γ(nα + α)
n=0
7 Nonhomogeneous equation
Consider the nonhomogeneous equation
Z t
g(t − s)u0 (s) ds = Au(t) + f (t), u(0) = 0 (7.1) nhom
0
2 (0, ∞, X) be a solution. Then, it is unique and satisfies the Laplace
Suppose u ∈ Lnhom
loc
transform of (7.1)
(∆(λ) I − A)û = fˆ. (7.2) lap
It follows from Theorem 2.3 that A generates the C0 semigroup T (t), t ≥ 0 on Z =
C((−∞]; X).
thm Theorem 7.1. For ω ≥ 0 we define the operator
and define Z t
u(t) = (A − ω I) T (t − s)Ψω f (s) ds (0). (7.3) nhom1
0
nhom1 nhom
Then, (7.3) defines the solution to (7.1) if f ∈ C 2 (0, T ; X).
Z t
z(t) = (A − ω I) T (t − s)Ψω f (s) ds
0
Z t Z t
0
= T (t)Ψω f (0) − Ψω f (t) + T (t − s)Ψω f (s) ds − ω T (t − s)Ψω f (s) ds
0 0
(7.4)
lap1
lap1
and thus, z ∈ C(0, T ; Z) and z(t) = u(t + ·). Taking the Laplace transform of (7.4),
we obtain
ẑ(λ) = (A − ω I)(λ I − A)−1 Ψω fˆ(λ)
26
res1
Since from (5.6) and
Z 0
∆(λ) = λ eλθ g(θ) dθ,
−∞
Z 0 Z 0
−1 −1 −1 ωθ
((λ I − A) Ψω )(0) = −(∆(λ) I − A) (∆(ω) I − A) g(θ)(e −λ eλ(θ−ξ) eωξ dξ)
−∞ θ
∆(ω) − ∆(λ)
= (∆(λ) I − A)−1 (∆(ω) I − A)−1 ,
λ−ω
Thus we have
û(λ) = (∆(λ) I − A)−1 fˆ(λ)
and the claim holds.
When 0 ∈ ρ(A), we let ω = 0 and obtain
Z t
−1 −1
u(t) = S(t)A f (0) − A f (t) + S(t − s)A−1 f 0 (s) ds
0
(7.5) nhom0
Z t Z t
= Ṡ(t − s)A−1 f (s) ds = P (t − s)f (s) ds,
0 0
we have
Z t
(∆(ω) I − A)y(t) − lim g0 (t − s)e−ω(t−s) (y(t) − y(s)) ds = e−ωt f (t)
→0+ −∞
27
thm6.2 Theorem
nhom
7.2. For f ∈ C(0, T ; X) nhom0
there exists a unique solution u in C(0, T ; X) to
(7.1) and it has the representation (7.5);
Z t
u(t) = P (t − s)f (s) ds.
0
and for some M >
|u(t)| ≤ M |f |C(0,t;X) .
7.1 f ∈ L2 (0, T ; H)
Let X = H be a Hilbert space. Since for R > T and 0 ≤ t ≤ T and C 1 (0, T ; H),
Z t Z 0
0
g(t − s)(u (s), u(s) − u(t)) ds = g(−θ)(u0 (t + θ), u(t + θ) − u(t)) dθ
t−R −R
1 0 0
Z
= ( g (−θ)|u(t + θ) − u(t)|2 dθ − g(R)|u(t) − u(0)|2 )
2 −R
and Z t Z 0
0
g(t − s)(u (s), u(s)) ds = g(−θ)(u0 (t + θ), u(t + θ)) dθ
t−R −R
d1 0
Z
= g(−θ)|u(t + θ)|2 dθ,
dt 2 −R
it follows from
Z T Z t Z TZ t Z T Z t
0 0
( g(t−s)u (s) ds, u(t)) dt g(t−s)(u (s), u(t)−u(s)) dsdt+ g(t−s)u0 (s), u(s)) dsdt
) 0 0 0 0 o
that Z 0 Z 0
2
g(θ)|u(T + θ)| dθ − g(θ)|x|2 dθ
−R −R
Z T Z 0
+ ( g 0 (θ)|u(t + θ) − u(t)|2 dθ + g(−R)|x − u(t)|2 ) dt (7.7) enrg
0 −R
Z T Z t
=2 g(t − s)u0 (s) ds, u(t)) dt.
0 0
Letting R → T + , we obtain
Z 0 Z 0
g(θ)|u(T + θ)|2 dθ − g(θ)|x|2 dθ
−T −T
Z T Z 0 Z T Z t
0
+ ( 2
g (θ)|u(t + θ) − u(t)| dθ + g(T )|x − u(t)| ) dt = 2 2
g(t − s)u0 (s) ds, u(t)) dt.
0 −t 0 0
(7.8) enrg1
Thus, we have
28
enrg1 Rt
cor6.1 Theorem 7.3. The energy identity (7.8) holds for all u ∈ L2 (0, T ; H) satisfying 0 g(t−
s)u0 (s) ds ∈ L2 (0, T ; H) and for all u(0) = 0.
Z T Z t Z T
0
( g(t − s)u (s) ds, u(t)) dt ≥ g(T ) | |u(t)|2 dt).
0 0 0
Z T Z 0 Z T
0 2 12 2
+ (− g (θ)|u(t + θ) − u(t)| dθ + g(T ) |x − u(t)| + ω |u(t)| ) dt ≤ |f (t)|2 dt.
0 −T ω 0
Otherwise, using exactly the same arguments as for the estimate for C(0, T ; H), we
have
Z 0
∆(ω) 2
g(θ)eωθ |y(T + θ)|2 dθ − |x|
−∞ ω
Z T Z 0 Z T Z 0
0 −ω(t−s)
− g (t − s)e 2
|y(s) − y(t)| ds dt + ω g(t − s)e−ω(t−s) |y(s)|2 ds dt
0 −∞ 0 −∞
Z T Z T
+ (∆(ω) I − 2A)y(t), y(t)) dt = 2 (e−ωt f (t), y(t)) dt
0 0
Consequently, we have
thm6.3 Theorem
nhom
7.4. For f ∈ L2 (0, T ; H) nhom0 u in L2 (0, T ; H) to
there exists a unique solutionenrg1
(7.1) and it has the representation (7.5) and the energy identity (7.8).
(Aφ, φ) ≤ −δ |φ|2V
nhom
Then, for f ∈ L2 (0, T ; H) nhom0
there exists a unique solution u in L2 (0, T ; V ) to (7.1) and
it has the representation (7.5) and the energy identity
Z t
1 t
Z
2
|u(s)|V ds ≤ |f (s)|2V ∗ ds.
0 δ 0
29
8 Dual System and Optimal Control Problems
fde
In this section we consider the dual system to (1.6) and its application to a class of
optimal control problems.
Let X = H be a Hilbert space. For u, p ∈ C 1 (0, T ; H) we have
Z T Z t Z T Z T
( g(t − s)u0 (s) ds, p(t)) dt = ( g(t − s)v(t) dt, u0 (s)) ds
0 0 0 s
Z T Z T Z T
d
= −(u(0), g(t)v(t) dt) − (u(s), g(t − s)p(t) dt) ds
0 0 ds s
Z T Z T Z T Z T
=( g(T − s)u(s) ds, p(T )) − (u(0), g(t)p(t) dt) − (u(s), g(t − s)p0 (t) dt) ds.
0 0 0 s
(8.1) dual
tm7 Theorem 8.1. Let Dtg : L2 (0, T ; H) → L2 (0, T ; H) be
Z t
g
(Dt u)(t) = g(t − s)u0 (s) ds
0
with domain
Then, Dtg is a densely defined, closed operator on L2 (0, T ; H) and the adjoint (Dtg )∗ is
given by Z T
d
((Dtg )∗ p)(s) = g(t − s)p(t) dt
ds s
with domain
Z T
dom ((Dtg )∗ ) 2
= {p ∈ L (0, T ; H) : g(t − s)p(t) dt ∈ H 1 (0, T ; X),
s
Z T
−
( g(t − s)p(t) dt)(T ) = 0}.
s
gs gs0
Proof: Since for g ∗ ∈ L1 (0, T ) is defined by (2.11) and from (2.12)
Z t
u(t) = g ∗ (t − s)(Dtg u) ds,
0
and Z T Z T Z T
(Dtg u, p(t)) dt = ( g(t − s)p(t) dt, u0 (s)) ds.
0 0 s
g
Since u ∈ dom(Dt ) are arbitrary,
Z T Z T
g(t − s)p(t) dt = y(t) dt, for a.e. s ∈ (0, T ).
s s
30
Hence, we obtain Z T
lim( g(t − s)p(t) dt) = 0
s↑T s
and Z T
d
y(s) = (Dtg )∗ p = g(t − s)p(t) dt.
ds s
Definition
fde
7.1 (Weak solution) A function u ∈ L2 (0, T ; H) is a weak solution to
(1.6) if
Z T Z T Z T
0 ∗
(− g(t − s)p (t) dt + A p(s), u(s)) + (p(s), f (s)) ds − (u(0), g(t)p(t) dt) = 0
0 s 0
(8.2) weak
for all p ∈ C 1 (0, T ; H) ∩ C(0, T ; dom(A∗ )) satisfying p(T ) = 0.
fde
Theorem 8.2. Weak solutio to(1.6) is unique
thm6.3
Proof: It follows from Theorem 7.4 that the dual system
Z T
g(t − s)p0 (t) dt = A∗ p(s) + f (s), p(T ) = 0 (8.3) duals
s
denote the control function and B ∈ L(U, X). The functional ` and h are convex on
X and U , respectively.
cost conts
cont Theorem 8.3. Problem (8.4)–(8.5) has an optimal control u∗ ∈ C.
conts
Proof: Since given u ∈ C (8.5) has a unique solution x(·; u) ∈ C(0, T ; X). Thus, the
optimal control problem is equivalent to minimizing
Z T
J(u) = (`(x(t; u)) + h(u(t))) dt
0
31
over u ∈ C. Suppose un ∈ C is a minimizing sequence of J over C, i.e. J(un ) ↓ and
limn→∞ J(un ) = η = inf u∈C J(u). Then there exists a weak convergent subsequence
(un , xn ) to (u∗ , x∗ ) in L2 (0, T ; U × X). Since
ofweak C is weakly closed, u∗ ∈ C. From
conts
(8.2) it follows that x is a weak solution of (8.5) corresponding to u∗ . Since convex
∗
32
9 Case:1 < α < 2
In this section we consider the case when 1 < α < 2;
Z t
α
Dt u = g2−α (t − s)u00 (s) ds = Au(t) + f (t), , u(0 = u0 , u0 (0) = v0 ,
0
or in general
Z t
g(t − s)u00 (s) ds = Au(t) + f (t), u(0 = u0 , u0 (0) = v0 . (9.1) a1-2
0
Equivalently, we have
Z t
d
u=v and g(t − s)v 0 (s) ds = Au(t) + f (t).
dt 0
with
u(0) = u0 , v(0) = v0 .
Assume −A is self-adjoint and positive on a Hilbert space H and define
1
V = dom((−A) 2 ) with |φ|2V = h−Aφ, φi.
A(u, z) = (z(0), z 0 )
with Z 0
dom(A) = {(u, z) ∈ X : A(u, z) ∈ X and g(θ)z 0 (θ) = Au}.
−∞
a12 Theorem 9.1. The linear operator A is m-dissipative and A generates a C0a1-2 semigroup
T (t) on X. u(t) = (T (t)(u0 , v0 ))1 ∈ C(0, T ; V ) defines a mild solution to (12.4).
Z 0
= g(θ)(z 0 (θ), z(θ) − z(0)) dθ.
−∞
sta0
From (2.1)
0
g(θ) − g(θ − ))
Z
(A(u, z), (u, z)) = − lim |z(θ) − z(0)|2 dθ,
→0+ −∞
and thus Z 0
(A(u, z), (u, z)) = − g 0 (θ)|z(θ) − z(0)|2 dθ ≤ 0.
−∞
For the resolvent
λ (u, z) − A(u, z) = (f 1 , f 2 )
33
is equivalent to
λ z − z0 = f 2, λ u − z(0) = f 1 .
From the first equation
Z 0
λθ
z(θ) = e z(0) + eλ(θ−ξ) f 2 (ξ) dξ (9.2) Res0
θ
and
Z 0 Z 0 Z 0
λ( eλθ g(θ) dθ)z(0) + g(θ)(λ eλ(ξ−θ) f 2 (ξ) dξ − f 2 (θ)) ds = Au.
−∞ −∞ θ
and
z(0) = λ u − f 1 . (9.4) Res2
Thus, R(λ I − A) = X.
Res1
It follows from (9.3) that u(t) = S1 (t)u + S2 (t)v with
Z γ+i∞
1
S1 (t)u = eλt (λ∆(λ) I − A)−1 ∆(λ)u dλ
2πi γ−i∞
(9.5) sol2
Z γ+i∞
1 ∆(λ)
S2 (t)v = eλt (λ∆(λ) I − A)−1 v dλ.
2πi γ−i∞ λ
Assume A is a sectorial operator, i.e., there exist M > 0, θ0 > 0 such that
M π
|(λ I − A)−1 | ≤ on Σθ0 = {λ ∈ C : arg(λ) ≤ + θ0 } ∩ {λ 6= 0}
|λ| 2
Assume that if λ ∈ Σθ0 , then λ∆(λ) ∈ Σθ̃ for θ̃ > 0. Let Γθ,δ be the integration path
defined by
π π
Γ± = {z ∈ C : |λ| ≥ δ, arg(λ) = ±( + λ)}, Γ0 = {z ∈ C : |z| = δ, |arg(z)| ≤ + θ}
2 2
for some δ > 0 and 0 < θ ≤ θ̃. Then, the solution map S1 (t) and S2 (t) is given by
Z
1
S1 (t)x = eλt (λ∆(λ) I − A)−1 ∆(λ)x dλ,
2πi Γ
Z
1
S2 (t)x = eλt (λ∆(λ) I − A)−1 ∆(λ)λ−1 x dλ,
2πi Γ
sol2
using the Cauchy integral representation (9.5) and the analytic continuation. Let
P (t) ∈ L(X), t > 0 be
Z
1
P (t)x = eλt (λ∆(λ) I − A)−1 x dλ.
2πi Γ
34
Then
d
S1 (t) = P (t)A
dt
a1-2
and the solution u(t) to (12.4) is given by
Z t
u(t) = S1 (t)u0 + S2 (t)v0 + P (t − s)f (s) ds
0
ukj−1 = uk−1
j , j ≤ 0, ukj = x for j ≤ −k.
That is,
ukj − uk−1
j ukj − ukj−1
= (10.2) dif1
h h
d
approximates u(t + ·) = Au(t + ·). Note that
dt
ukj = uk+j
0 = uk+j .
dif
Thus, (10.1) is equivalent to
−(k−1)
uk − uk−1 X uk+j − uk+j−1
g0 + gj = Auk + f k (10.3) dif0
h h
j=−1
fde
which is an approximation for (1.6) directly. We have the following stability results for
dif
(10.1);
stab1 Theorem 10.1. If A is dissipative, for f = 0 we have |uk0 | ≤ |x|. In general we assume
that for all u ∈ dom(A), there u∗ ∈ F (u) such that
hAu, u∗ i ≤ −δ |u|2 , (10.4) cond1
then
1
max |uk | ≤ max |f k |.
0≤k≤N δ 0≤k≤N
35
Proof:For f = 0 since
−(k−1) −(k−2)
X uj − uj−1 X gj−1 − gj
gj =− (uj−1 − u0 ), (10.5) dif1
h h
j=0 j=0
dif
it follows from (10.1) that for u∗ ∈ F (uk0 )
−(k−2)
X gj−1 − gj
hAuk0 , u∗ i = − ((ukj−1 , u∗ ) − |uk0 |2 ).
h
j=0
−(k−2)
X gj − gj−1
hAu0 , u∗ i > (|u0 |2 − |ukj−1 ||u0 |) > 0
h
j=0
which contradicts to the fact that A is dissipative. Thus, maxj≤0 |ukj | ≤ maxj≤0 |uk−1
j |.
k j
dif1
In general, suppose |u | ≥ |u | for 0 ≤ j ≤ N . From (10.5)
For all k ≥ 1
−(k−1) k −(`−2) k −(k−1) k
X X X gj − gj−1 X X 1X k2
gj |ukj |2 + |u`j −u`0 |2 h+δ |u`0 |2 h ≤ gj |x|2 + |f | h.
h δ
j=0 `=1 j=0 `=0 j=0 `=1
(10.6) est2
dif1
Proof: From (10.5)
36
dif
Since from (10.1)
X ukj − ukj−1 k
gj ( , u0 ) = (Auk0 + f k , uk0 ),
h
j≤0
summing this over k we obtain
−(k−1) k −(`−2) −(k−1) k
X X X gj − gj−1 X X
gj |ukj |2 + |u`j − u`0 |2 h = gj |x|2 + 2 (Au`0 + f ` , u`0 ) h.
h
j=0 `=1 j=0 j=0 `=1
where Z 0
k
E = g(θ)(u0 (kh + θ) − U 0 (khθ)) dθ → 0
−kh
Let X = H be a Hilbert space. The, we have
N
X
|E ` |2 h → 0 as h → 0+
`
If we define
t − kh k
uh (t) = uk + (u − uk−1 ) if t ∈ (k − 1)h, kh]
h
est2 dif2
then it follows from (10.6) and (10.7) that
1
|uh − Uh |L2 (0,T ;H) → |Eh |L2 (0,T ;H) → 0
δ
as h → 0+ .
In general, let X be a Banach space and then we have;
Corollary 10.1.
|uh − Uh |C(0,T ;X) → 0, as h → 0+ ,
dif2
assuming (10.7) and f ∈ C(0, T ; X).
dif2 form
Proof: Using the same arguments as above, from (10.7) and (10.8) we have
1
|uh − Uh |C(0,T ;X) ≤ |Eh |C(0,T ;X) → 0.
δ
37
10.1 Cone invariance and Maximum Principle
Let H = L2 (Ω) and C be a a closed cone in H. A is cone preserving, i.e.,
(I − s A)−1 C ⊂ C
11 Semi-linear equations
In this section we consider the semilinear equation
or equivalently Z t
x(t) = S(t)x0 + P (t − s)F (x(s)) ds, t ≥ 0 (11.1) Lip
0
with the locally Lipschitz function F in X. Assume
τα
M0 (1 + τ α ) + CM ρ(M ) ≤ M, (11.4) tau
α
38
cond
given M0 > 0. It follows from (11.5)
τα
|(Ψx1 )(t) − (Ψx2 )(t)| ≤ Cρ(M ) |x1 − x2 |C(0,τ ;X)
α
for |x1 |, |x2 | ≤ M . Thus, Ψ has
tau
a unique fixed point x ∈ C(0, τ ; X) satisfying |x(t)| ≤
α
M, t ∈ [0, τ ], provided that (11.4) holds, since Cρ(M ) τα < 1.
Hence, given t ≥ 0 for h ≥ 0
Z t Z t+h
x(t + h) = S(t + h)x0 + P (t + h − s)F (x(s)) + P (t + h − s)F (x(s)) ds
0 t
Z h
= f (t + h) + P (h − σ)F (x(t + σ)) dσ
0
fixed
has a unique solution x(t + h), 0 ≤ h ≤ τ as a fixed point to (11.3) with
Z t
f (t + h) = S(t + h)x0 + P (t + h − s)F (x(s)) ds, ; h ≥ 0
0
1
≤ CM γ [(t − 1)α − tα + t−α(γ−1) ].
Γ(1 − α)Γ(α)
For t ≥ 2
Z t Z t−1 Z 1
−1+α γ −γα −1 γ −γα
I(t) ≤ C(t − s) M s ) ds + C(t − s) M s ds + C(t − s)−1 M γ ds
t−1 1 0
≤ CM γ [(t − 1)−γα tα + 1
Γ(δ)Γ(1−δ) M
γ + (t − 1)−1 tα ]t−α
with δ = α(γ − 1). It thus follows that there exists β independent of t ≥ 0 such that
39
It can be proved that if M0 > 0 is sufficiently small there exists M > 0 such that
cond
TheoremLip
(Asymptotic Stability) Assume ρ(s)s = sγ and condition (11.5) holds.
Then, (11.1) has a unique solution in C(0, ∞; X) satisfying |x(t)| ≤ M min(1, t−α ).
Proof: Using exactly the same arguments as above, without loss of the generality with
same β > 0 we have
|Ψx1 − Ψx2 | ≤ βCM γ−1 |x1 − x2 |.
The claim follows since βCM γ−1 < 1.
|A−α1 (F (x) − F (y))| ≤ ρ(M ) |Aα2 (x − y)| for |Aα2 x|, |Aα2 y| ≤ M. (11.6) cond1
and
|A−α1 F (0)|X ≤ c. (11.7) cond2
with 0 ≤ α1 + α2 < 1. Let Xα2 = dom(Aα2 ). For u ∈ C([0, τ ], Xα2 ) define the map
Z t
(Ψ(u))(t) = S(t)x0 + P (t − s)F (u(s)) ds.
0
Lip
Theorem 11.1. (Local Existence) For x0 ∈ Xα2 there exists a τ > 0 such (11.1)has a
unique solution u ∈ C(0, τ ; Xα2 ).
cor5.4
Proof: For u(¸0, τ, Xα2 ) satisfying |u| ≤ M on [0, τ ] it follows from Corollary 5.4 that
Rt
|Aα2 Ψ(u)| = |S(t)||Aα2 x0 | + | 0 Aα1 +α2 P (t − s)A−α1 F (u(s)) ds|.
Z t
α2
≤ |A x| + C(t − s)−β (ρ(M )M + c) ds
0
C
≤ |Aα2 x| + ( τ 1−β (ρ(M )M + c)
1−β
where β = α (1 − (α1 + α2 )). Let τ = τ (M0 ) such that
C
M0 + τ 1−β (ρ(M )M + c) ≤ M
1−β
cond1
given M0 = |Aα2 x0 |. It follows from (11.6) that
C
|Aα2 (Ψ(u1 ) − Ψ(u2 ))| ≤ τ 1−β ρ(M ) |Aα2 (u1 − u2 )|
1−β
Thus, Ψ has a unique fixed point
Lip
in x ∈ C(0, τ, Xα2 ) satisfying |Aα2 x(t)| ≤ M on [0, τ ],
which defines a solution to (11.1)
40
12 Examples
In this section we discuss the application of our theory for concrete examples.
Au = ∇(a(x)∇u) + b(x) · ∇u
with a ∈ Rd×d ∈ C 1 (Ω) is symmetric and positive definite and b ∈ C(Ω). The linear
operator A with
dom (A) = {u ∈ C 2 (Ω) ∩ H01 (Ω)}
is dissipative in X = C(Ω). In fact, if u(ζ0 ) ≥ |u| for ζ0 ∈ Ω, then (∇u)(ζ0 ) = 0 and
H(x0 ) ≤ 0. and
where Hi,j = uxi ,xj is the Hessian of u. Similarly, if u(ζ0 ) ≤ |u|, then (Au)(ζ0 ) ≥ 0.
Define the nonlinear operator by
Au = −(f (u))x ,
Aφ = φ0 (θ)
with domain Z 0
dom(A) = { g(θ)φ0 (θ) dθ = Aφ(0)}.
−∞
41
Let
C = {φ ∈ Z : φ ≥ 0}.
Since Ac = 0 for all constant c, it follows that
φ − c ∈ C =⇒ (I − λ A)−1 φ − c ∈ C.
Similarly,
c − φ ∈ C =⇒ c − (I − λ A)−1 φ ∈ C.
Thus, without loss of generality, one can assume f is bounded. Let ρ ∈ C 2 (R) be
a monotonically increasing function satisfying ρ(0) = 0 and ρ(x) = sgn(x), |x| ≥ 1.
Note that
−(f (u1 )x − f (u2 )x , ρ(u1 − u2 )) = (f (u1 ) − f (u2 ), ρ0 (u1 − u2 ) (u1 − u2 )x ),
where Z 1
η= fu (u2 + τ (u1 − u2 )) dτ.
0
Rx
If we define Ψ(x) = 0 σρ0 (σ) dσ, then
u1 − u2
|(η (u1 − u2 ), ρ0 (u1 − u2 ) (u1 − u2 )x )| = (Ψ( ), ηx ) ≤ const |ηx |1 → 0
as → 0. Note that for u ∈ L1 (Rd )
(u, ρ (u)) → |u| and (ψ, ρ (u)) → (ψ, sgn0 (u)) for ψ ∈ L1 (Rd )
as → 0+ . Thus,
hAu1 − Au2 , sgn0 (u1 − u2 )i ≤ 0
IK
and A is monotone. It is show in [2] that
range(λ I − A) = X,
for all ψ ∈ C01 (Rd ) and k ∈ R. Hence A has a maximal monotone extension in L1 (Rd ).
42
12.3 Fractional Hamilton-Jacobi equation
Rx
Let u is a solution to a scalar conservation in R1 , then v = u dx satisfies the fractional
Hamilton-Jacobi equation
Dtα v + f (vx ) = 0.
Let X = C0 (Rd ) and
That is, there exists a unique viscosity solution to λv − f (vx ) = g; for all φ ∈ C 1 (Ω) if
v − φ attains a local maximum at x0 ∈ Rd , then
Dtα u + u · ∇u + grad p = ν ∆u
(12.5) NS
div u = 0
where u is the velocity field defined on domain Ω with Lipschitz boundary Γ and p is
the pressure. Let V be the divergence free closed subspace of H01 (Ω)d :
43
Let P be the orthogonal projection of L2 (Ω)d onto X and define the Stokes operator
−A by
−Au = P ∆u with dom (A) = H 2 (Ω)d ∩ V
and the convection
F (u) = −P (u · ∇u) for u ∈ V
NS
Then, (12.5) is equivalently written as
Dtα u = Au + F (u).
A is maximal
Using exactly the same arguments as for Theorem 3.1, fdes1 monotone in Z
fdes2
and generate a nonlinear semigroup on Z. Thus Both (13.1) and (13.2) has a unique
mild solution u ∈ C(0, T ; X).
44
14 Fractional equations in Space
In this section we consider the nonlocal diffusion equation of the form
Z
ut = Au = J(z)(u(x + z) − u(x)) dz.
Rd
Or, equivalently
Z
Au = J(z)(u(x + z) − 2u(x) + u(x − z)) dz
(R)d )+
Thus, Z
(Au, φ)L2 = J(z)∇u(x + z) · ∇φ(x) dz dx
Rd ×Rd
Under the kernel J is completely monotone, one can prove that A is a maximal mono-
tone extension.
(λ, u − ψ) = 0, λ ≤ 0, u≥ψ
where the generator B for the jump process is given by
Z ∞
Bu = k(s)(u(x + s) − u(x) + (es − 1)ux ) ds.
−∞
Since
Z ∞ Z ∞ Z ∞
k(s)(u(x + s) − u(x) ds = +
k (s)(u(x + s) − u(x)) ds + k − (s)(u(x − s) − u(x)) ds
−∞ 0 0
∞ ∞
k + (s) + k − (s) k + (s) − k − (s)
Z Z
= (u(x + s) − 2u(x) + u(x − s)) ds + (u(x + s) − u(x − s)) ds.
0 2 0 2
45
Thus, Z ∞ Z ∞
( k(s)(s)(u(x + s) − u(x) ds)φ dx
−∞ −∞
Z ∞ Z ∞
= ks (s)(u(x + s) − u(x))(φ(x + s) − φ(s)) ds dx
−∞ 0
Z ∞ Z ∞
+ ( ku (s)(u(x + s) − u(s)))φ(x) dx
−∞ 0
where
k + (s) + k − (s) k + (s) − k − (s)
ks (s) = , ku (s) =
2 2
and hence
Z ∞ Z ∞
(Bu, φ) = ks (s)(u(x + s) − u(x))(φ(x + s) − φ(s)) ds dx
−∞ −∞
Z ∞ Z ∞ Z ∞
+ ( ku (s)(u(x + s) − u(s)))φ(x) dx + ω ux φ dx.
−∞ −∞ −∞
where Z ∞
ω= (es − 1)k(s) ds.
−∞
If we equip V = H 1 (R) by
Z ∞Z ∞
σ2 ∞
Z
|u|2V = ks (s)|u(x + s) − u(x)|2 ds dx + |ux |2 dx,
−∞ −∞ 2 −∞
where we have
u(x + s) − 2u(x) + u(x − s) s2
∼ uxx (x) + uxxxx (x) + O(s4 )
s2 12
We apply the fourth order approximation of uxx by
u(x + h) − 2u(x) + u(x − h) 1 u(x + 2h) − 4u(x) + 6u(x) − 4u(x − h) + u(x − 2h)
uxx (x) ∼ −
h2 12 h2
46
and we apply the second order approximation of uxxxx (x) by
by
uk+1 − 2uk + uk−1 1 uk+2 − 4uk+1 + 6uk − 4uk−1 + uk−2
ρ0 ( 2
− )
h 12 h2
ρ1 uk+2 − 4uk+1 + 6uk − 4uk−1 + uk−2
+ ,
12 h2
where Z h Z h
2
2 1 2
ρ0 = s ks (s) ds and ρ1 = 2 s4 ks (s) ds.
−h h −h
2 2
s2 00
u(xk+j + s) ∼ u(xk+j ) + u0 (xk+j )s + u (xk+j )
2
with
uk+j+1 − uk+j−1
u0 (xk+j ) ∼
2h
uk+j+1 − 2uk+j + uk+j−1
u00 (xk+j ) ∼ .
h2
That is,
Z (k+ 12 )h
u(xk+j + s)ks (s) ds
(k− 12 )h
1
R (k+ 12 )h
ρk1 = h (k− 12 )h
(s − xk )ks (s) ds
1
R (k+ 12 )h
ρk2 = h2 (k− 21 )h
(s − xk )2 ks (s) ds.
47
For the skew-symmetric integral
Z h
2 ρ3
ku (s)(u(x + s) − u(x − s)) ds ∼ ρ2 ux (x) + h2 uxxx (x)
− h 6
2
where Z h Z h
2 1 2
ρ2 = 2sku (s) ds, ρ3 = 2s3 ku (s) ds.
−h h2 −h
2 2
uk+1 − uk−1 uk+2 − 2uk+1 + 2uk−1 − uk−1 ρ3 uk+2 − 2uk+1 + 2uk−1 − uk−1
∼ ρ2 ( − )+ .
2h 6h 6 h
with
dom (A) = {u ∈ H 1 (0, 1) : u0 ∈ dom(Dtα ) with u(0) = u(1) = 0}.
Since
(t − s)−α 00 (t − s)−α
Z 1 Z t Z 1 Z 1
( u (s) ds)φ(t) dt = ( φ(t) dt)u00 (s) ds
0 0 Γ(1 − α) 0 s Γ(1 − α)
1
(t − s)−α 1
t−α 1 1
(t − s)−α
Z Z Z Z
0 d
= u (1)( φ(t) dt)(1− ) − u0 (0) φ(t) dt − 0
u (s) φ(t) dt
s Γ(1 − α) 0 Γ(1 − α) 0 ds s Γ(1 − α)
1
(t − s)−α 1
t−α 1
d2 1
(t − s)−α
Z Z Z Z
0
= u (1)( φ(t) dt)(1− ) − u0 (0) φ(t) dt + u(s) 2 φ(t) dt ds,
s Γ(1 − α) 0 Γ(1 − α) 0 ds s Γ(1 − α)
the adjoint operator of A is given by
(t − s)−α 0
Z 1
∗ d
A φ=− φ (t) dt + q(s)φ(s)
ds s Γ(1 − α)
48
with
1
(t − s)−α
Z
∗ 2
dom (A ) = {φ ∈ L (0, 1) : φ(t) dt ∈ H 2 (0, 1) ∩ H01 (0, 1)}
s Γ(1 − α)
1
t−α 1
(t − s)−α 0
Z Z
1
= {φ ∈ H (0, 1) : φ(1) = 0; φ(t) dt = 0, φ (t) dt ∈ H 1 (0, 1)}
0 Γ(1 − α) s Γ(1 − α)
Thus, dom (A) 6= dom (A∗ ) and A is not self-adjoint. Note that
Z 1
(φ, A∗ φ) = φ0 (s)Ds1 φ ds.
0
where Lk (·) is the k-th Legendre polynomial on [−1, 1]. The boundary condition
uN (0) = uN (1) = 0 implies
X X
uN = − uk , uN −1 = uk (15.1) Leg
k:even k:odd
if N is even and X X
uN = − uk , uN −1 = uk (15.2) Leg2
k:odd k:even
−2
if N is odd. Thus, {uk }N
k=0 defines the approximations
N
X −2
N
u = uk Lk (2t − 1)
k=0
and
ũN = uN + uN −1 LN −1 (2t − 1) + uN LN (2t − 1).
Leg Leg2
where uN −1 and uN are defined by (15.1)–(15.2). The Legendre-tau approximation
AN : X n−2 → X N −2 is defined by
AN uN = P N −2 AũN ,
49
That is,
1 t
(t − s)−α N 0
Z Z
N N
(A u )j = Lj (2t − 1) − (ũ ) (s) ds + q(t)uN (t) dt.
0 0 Γ(1 − α)
where we let X X
βN = − βk , βN −1 = − βk
k:even k:odd
and
N −2
1
(t − s)−α N
Z X
N −2
P φ (t) dt = βk Lk (2s − 1).
s Γ(1 − α)
k=0
(t − s)−α N
Z 1
d2
∗ N N N −2
(A ) φ = 2 P φ (t) dt + βN −1 LN −1 (2s − 1) + βN LN (2s − 1) +P N −2 (q(s)φN ).
ds s Γ(1 − α)
16 Appendix: Derivation
Let us consider the continuous time random walk (CTRW) process x(t) on R which
is characterized by the joint probability density function (pdf) ϕ(ξ, τ ) of jumps ξ =
x(tj ) − x(tj−1 ) and waiting times τj = tj − tj−1 . We assume jumps and waiting times
are independent, i.e., ϕ = λ(ξ)ψ(τ ). The jump pdf λ(ξ) represents the pdf of jump
size ξ and the waiting time pdf ψ(τ ) represents the pdf of waiting time τ . Thus,
Z t
ψ(τ ) dτ
0
gives the probability tat at least one jump is taken in (0, t).
Z t
Ψ(t) = 1 − ψ(τ ) dτ
0
defines the probability of no jump occurs during (0, t]. Let us denote by p(x, t) the
pdf of reaching to position x after time t, i.e. p(x, 0) = δ(x). The master equation of
CTRW is given by
Z t Z ∞
p(x, t) = δ(x)Ψ(t) + ψ(t − t0 ) λ(x − x0 )p(x0 , t0 ) dx0 dt0 . (16.1) CTRW
0 −∞
CTRW
Taking the Laplace transform in t and the Fourier transform in x of (16.1), we obtain
the Montroll-Weiss equation
Ψ̂(s) 1 − ψ̂(s) 1
p̂(k, s) = = . (16.2) MW
1 − λ̂(k)ψ̂(s) s 1 − λ̂(k)ψ̂(s)
50
InMWorder to derive an evolution equation of Fokker-Planck-Kolmogorov type we rewrite
(16.2) as
where
Ψ̂(s)
Φ(t) = L−1 { }
1 − sΨ̂(s)
and Z t
Ψ(t) = Φ(t − s)ψ(s) ds.
0
If Φ̂(s) = 1 and thus ψ(t) = Ψ(t) = e−t , it reduces to the Kolmogorov-Feller equation
Z ∞
∂
p(x, t) = −p(x, t) + λ(x − x0 )p(x0 , t) dx0 .
∂t −∞
If we assume λ is symmetric,
Z ∞ Z ∞
−p(x, t) + λ(x − x0 )p(x0 , t) dx0 = λ(s)(p(x + s) − 2p(x) + p(x − s)) ds
−∞ 0
and thus
Z ∞ Z ∞
[ λ(s)(p(x + s) − 2p(x) + p(x − s)) dsψ(x) dx
−∞ 0
Z ∞ Z ∞
=− λ(s) (p(x + s) − p(x))(ψ(x + s) − ψ(x)) dx.
0 −∞
Hence, the right hand side defines a self-adjoint nonnegative definite operator A on
L2 (R).
51
Consider a diffusive mass transport of chemical specifies through a rigid porous
saturated composite, consisting of tow porous materials;
∂c x x
= ∇ · (D( )∇c + b( ) c) (17.1) model
∂t
where we assume the periodic diffusive media D( x ) and the periodic advection b( x )
with period and they are given by
D(y) = 2 D2 χΩ2 (y) + D1 χΩ1 (y), b(y) = 2 b2 χΩ2 (y) + b1 χΩ1 (y),
Here subdomains Ω1 and Ω2 are for each composite and are disjoint and
Ω1 ∪ Ω2 = [0, 1]d = Ω. Γ = Ω1 ∩ Ω2 .
Here > 0 is a small parameter which measures the characteristic length of the period
of the heterogeneities compared to a macroscopic length. the heterogeneity reflects
that Ω1 is a diffusive (fluid) medium and Ω2 is a less diffusive (solid) medium with
ration 2 . It follows from [Auriault&Lewandowka] that the homogenized equation as
→+ is given by
Z t
|Ω1 ∂c ∂c
+ K(t − s) ds = ∇ · (Deff ∇c + beff c), (17.2) hom0
|Ω| ∂t 0 ∂s
where Z
1
L(K) = h1 − k̂i = (1 − k̂) dy
|Ω| Ω2
−1 order:
∇y · (D1 ∇x ĉ01 + b1 ĉ01 ) + ∇x · (D1 ∇y ĉ01 ) + ∇y · (D1 ∇x ĉ11 ) = 0, y ∈ Ω1
(17.4) order-1
n · (D1 ∇y ĉ11 + D1 ∇x ĉ01 + b1 ĉ01 ) = 0, y ∈ Γ
52
0 order:
∇x · (D1 ∇x ĉ01 + b1 ĉ01 ) + ∇x · (D1 ∇y ĉ11 ) + ∇y · (D1 ∇x ĉ11 + b1 ĉ11 ) + ∇y · (D1 ∇y ĉ21 ) = pĉ01 , y ∈ Ω1
and thus
ĉ02 = (1 − k̂)ĉ, (17.10) homo2
where Y −periodic function k̂ satisfies
53
where
∂wk
Z
1
(Deff )kj = D1 (Ikj + ) dy
|Ω| Ω1 ∂yj
Z (17.11) homo3
1
beff = (b1 + D1 ∇v) dy.
|Ω| Ω1
order0-2 homo2
From (17.6) and (17.10)
Z Z
1 1
n · (D2 ĉ02 ) ds = − p(1 − k̂) dy ĉ
|Ω| Γ |Ω| Ω1
Hence we obtain
|Ω1 |
Z
1
∇x (Deff ∇x ĉ + beff ĉ) + (1 − k̂) dy pĉ = pĉ.
|Ω| Ω1 |Ω|
hom0
which implies (17.2).
If û = 1 − k̂ we have
D2 ∆û = p û
(17.12) hom
û = 1 at Γ
For Ω2 = {|r| ≤ r0 } in R3 the radial solution û = û(r) satisfies
(ru)00 = pr u(r)
Thus, √
r0 sinh( pr)
û(r) = √
r sinh( pr0 )
and
4πr2 √ 4πr0
hûi = √ 0 coth( pr0 ) − .
p p
hom
In general, from (17.12) by the divergence theory
Z
∂
phûi = û ds.
Γ ∂ν
Since at x∗ ∈ Γ
∂2
û + ∆τ û = pû
∂ν 2
and û = 1 on Γ, we have
∂2 ∂
2
û − κ û = pû,
∂ν ∂ν
where κ is the curvature of Γ at x∗ . It can be shown that
∂ √
û ∼ p
∂ν
for p >> 1, since √
p (x−x∗ )·ν
û ∼ e at x∗ ∈ Γ.
54
Thus, for p >> 1
|Γ| 1
hûi ∼ √
|Ω| p
and 1
|Γ| t− 2
K(t) ∼ .
|Ω| Γ( 12 )
hom0
It thus follows from (17.2) that
|Ω1 | ∂ |Γ| 12
c+ D c = ∇ · (Deff ∇c + beff c).
|Ω| ∂t |Ω| t
References
B [1] E. Bazhrlkova, Fractional Evolution Equations in Banach spaces, University Press
Facilities, Eindhoven, University Technology (2001).
IK [2] K. Ito and F. Kappel, Evolution Equations and Approximations, World Scientific,
(2002).
MK [3] R. Metzler and J. Klafter, The restaurant at the end of the random walk: recent
developments in the description of anomalous transport by fractional dynamics. J.
Physics A 37 (2004), R161R208.
P [4] J Pruss, Evolutionary Integral Equations and Applications, Birkhauser, Basel,
(1993).
SGM [5] E. Scalas, R. Gorenflo and F. Mainardi, Fractional callculus and continoous-time
finance, Physica A, 284 (2000), 376-384.
55