Documente Academic
Documente Profesional
Documente Cultură
DYNAMICAL SYSTEMS
Supplement 2013
Website: www.aimSciences.org
pp. 259272
Matthew A. Fury
Division of Science and Engineering
Penn State Abington
1600 Woodland Road
Abington, PA 19001, USA
259
260
MATTHEW A. FURY
approximating a known solution of an ill-posed problem by the solution of an approximate well-posed problem
dv
= f (A)v(t)
dt
v(0) = ,
0t<T
(2)
where > 0 and the operator f (A) approaches A in some sense as 0. For
instance, introduced by Lattes and Lions [12] and Miller [15], the quasi-reversibility
method yields a common approach in the regularization of problem (1) where
f (A) = A A2 (cf. [14, Chapter 3.1.1]). Specifically, regularization refers to
the following:
Definition 1.1. ([10, Definition 3.1]) A family {R (t) | > 0, t [0, T ]} of
bounded linear operators on H is called a family of regularizing operators for the
problem (1) if for each solution u(t) of (1) with initial data H, and for any
> 0, there exists () > 0 such that
(i) () 0 as 0, and
(ii) ku(t) R() (t) k 0 as 0 for 0 t T whenever k k .
Hence, in Definition 1.1, the solution u of (1) may be approximated given a small
perturbation in the operator (A to f (A)) and also in the initial data ( to ).
Here the regularizing operator R (t) may be defined as the solution operator of
problem (2); that is R (t)x is the solution of (2) with initial condition v(0) = x.
Regularization has been established by authors such as Melnikova and Filinkov
[14], Melnikova [13], Ames and Hughes [3], Huang and Zheng [9, 10], Trong and
Tuan [19, 20], and Fury and Hughes [7] for evolution problems in various settings.
For example, Melnikova, Ames and Hughes, and Huang and Zheng each prove regularization for problem (1) in the case that A generates a holomorphic semigroup
on a Banach space (cf. [13], [3], [9, 10]). Also, Fury and Hughes prove regularization
for nonautonomous evolution problems where the operator A in (1) is replaced by
the nonconstant operator A(t) for t [0, T ] (cf. [7]).
This paper extends regularization to evolution problems that are both nonautonomous and inhomogeneous in a Hilbert space H, problems specifically of the
form
du
= A(t, D)u(t) + h(t) 0 s t < T
(3)
dt
u(s) =
Pk
j
where D is a positive, self-adjoint operator in H, A(t, D) =
j=1 aj (t)D for
+
t [0, T ] with aj C([0, T ] : R ) for 1 j k, and h is a function from [s, T ]
into H. Since D is positive, self-adjoint and aj (t) is nonnegative for 1 j k,
(3) is generally ill-posed. Taking the simple example where A(t, D) = D =
for t [0, T ], H = L2 (Rn ), and h 0, (3) reduces to the ill-posed backward heat
equation.
Following techniques of Lattes and Lions [12], Miller [15], Ames [2], Showalter
[18], and Ames and Hughes [3], we define the approximate well-posed problem
dv
= f (t, D)v(t) + h(t)
dt
v(s) =
0st<T
(4)
261
where 0 < < 1. We will show that the assumptions made on the operators
f (t, D) yield an evolution system V (t, s), 0 s t T (cf. [16, Definition 5.1.3]) associated with problem (4) such that the family of bounded operators
{V (t, s) | > 0, 0 s t T } regularizes the ill-posed problem (3) in the
sense of regularization for inhomogeneous problems. That is given a small change
in initial data k k , there exists () > 0 such that 0 as 0 and
Z t
V (t, r)h(r)dr)k 0 as 0
ku(t) (V (t, s) +
s
ts
ku(t) v (t)k C T s M T s ,
0stT
(5)
where v (t) is the solution of (4). Allowing to approach 0, the inequality (5)
establishes H
older-continuous dependence on modeling (cf. [12], [15], [2], [18], [3],
[4], [6], [7]) meaning a small change from the models (3) to (4) produces only a
small change in the corresponding solutions. The calculations and results in this
section of the paper extend the work of Fury and Hughes who in [6] prove Holdercontinuous dependence on modeling for the homogeneous version of (3) where h 0.
In Section 4, we then use inequality (5) to prove the main result of the paper which is
the existence of a family of regularizing operators for problem (3). Finally, Section 5
applies the theory to partial differential equations in L2 (Rn ).
Below, B(H) will denote the space of bounded linear operators on H. For a
subspace Y of H, B(Y, H) denotes the space of bounded linear operators from Y
into H, and k kY H the corresponding operator norm. When referring to solutions
of (3) and (4), we will be concerned with classical solutions which (in the case of
(3)) are functions u : [s, T ] H such that u(t) Dom(A(t, D)) for s < t < T ,
u C([s, T ] : H)C 1 ((s, T ) : H), and u satisfies (3) in H. Also, a Y -valued solution
of (4) is a classical solution v of (4) such that v C([s, T ] : Y ).
2. The approximate well-posed problem. For 0 < < 1, let f : [0, T ]
[0, ) R be a function continuous in t [0, T ] and Borel in [0, ). In this
section, we first determine conditions on the function f so that the approximate
problem (4) is well-posed, meaning a unique classical solution exists for each initial
262
MATTHEW A. FURY
for x Dom(f (t, D)) = {x H | 0 |f (t, )|2 d(E()x, x) < }, where {E()}
denotes the resolution of the identity for the self-adjoint operator D.
Proposition 2.1. Let H be a Hilbert space and for 0 < < 1, let f : [0, T ]
[0, ) R be continuous in t [0, T ] and Borel in [0, ). Assume the
following conditions.
(i) There exists R such that f (t, ) for all (t, ) [0, T ] [0, ).
(ii) There exists a Borel function r : [0, ) [0, ) such that |f (t, )| r ()
for all (t, ) [0, T ] [0, ), and Dom(f (t, D)) = Dom(r (D)) for all t [0, T ].
(iii) The function t 7 f (t, D) is continuous in the B(Y, H) norm k kY H where
Y = Dom(r (D)) and k kY denotes the graph norm associated with the operator
r (D).
(iv) h C([s, T ] : Y ).
Then
(4) is well-posed
with unique classical solution (and Y -valued solution) v (t) =
Rt
Rt Rt
e s f (,D)d + s e r f (,D)d h(r)dr for every Y .
Proof. As seen in [6, Proposition 2.10], condition (i) implies that for each t [0, T ],
f (t, D) is the infinitesimal generator of the C0 semigroup esf (t,D) , s 0, and
conditions (i)(iii) imply that {f (t, D)}t[0,T ] is a stable family satisfying Katos
stability conditions (cf. [11, Theorem 4.1 (i)(iii)], [16, Section 5.3]) with Y =
Dom(r (D)) and with stability constants K = 1 and . The homogeneous
version
Rt
f (,D)d
s
of (4) with h 0 is then well-posed with
unique
Y
-valued
solution
t
7
e
Rt
f
(,D)d
263
In light of Condition (A, p), for t [0, T ], we set g (t, ) = A(t, ) + f (t, ).
For each n | |, set en = { [0, ) | max{|g (t, )| : t [0, T ]} n}. Then
for en , max{|A(t, )| : t [0, T ]} n + and max{|f (t, )| : t [0, T ]}
2n + . Set En = E(en ) and let n En H. Consider the evolution problem
du
= A(t, D)En u(t) + hn (t)
dt
u(s) = n
0st<T
(6)
264
MATTHEW A. FURY
that g (t, ) for all (t, ) [0, T ] [0, ). Then if u(t) and v (t) are classical
solutions of (3) and (4) respectively with Y and h R C([s, T ] : Y ), and if
T
there exist constants
M 0 , M 00 0 such that kp(D)b(D)e s A(,D)d k M 0 and
RT
kp(D)b(D)e s A(,D)d h(t)k M 00 for all t [s, T ], then there exist constants C
and M independent of such that for 0 s t < T ,
T t
ts
ku(t) v (t)k C T s M T s .
Proof. Let n = En for n | |, and define S to be the complex strip S =
{t + i | t [s, T ], R}. Since D is self-adjoint, eiD is a bounded operator on
H for every R. Hence, for = t + i S, we may define n : S H by
n () = eiD (un (t) v,n (t))
Rt
where, as above, un (t) = Un (t, s)n + s Un (t, r)hn (r)dr and v,n (t) = V,n (t, s)n +
Rt
V (t, r)hn (r)dr.
s ,n
The function n () is not necessarily analytic, and so we may not apply the
Three Lines Theorem directly to it. To construct a function that is analytic on the
interior of S, following work of Agmon and Nirenberg [1, p. 148], we consider the
function
Z Z
1
1
1
n (z)
+
dxdy,
n () =
z z + 1 +
S
where = t + i and z =
x + iy are in S, and denotes the Cauchy-Riemann
1
n (z). Since
operator = 2 x + i y . Let z = x + iy S. We first determine
eiyD is a bounded operator on H, and since un (x) and v,n (x) are both in En H,
we have by Lemma 2.3,
d
d
iyD
n (z) =
e (un (x) v,n (x)) = eiyD (iD)(un (x) v,n (x)),
y
y
and therefore,
n (z)
2
n (z) + i n (z)
x
y
=eiyD (A(x, D)un (x) f (x, D)v,n (x)) eiyD (Dun (x) Dv,n (x)).
=
(7)
n (z) is
Following the method of Agmon and Nirenberg [1, p. 148], we show that
iyD
bounded and continuous on S. Since ke k = 1, we have from (7),
n (z)k kA(x, D)un (x) A(x, D)v,n (x)k
2k
+ kA(x, D)v,n (x) f (x, D)v,n (x)k
+ kDun (x) Dv,n (x)k.
(8)
265
=kA(x, D)Un (x, s)n A(x, D)W,n (x, s)Un (x, s)n k
Z x
kA(x, D)Un (x, r)hn (r) A(x, D)W,n (x, r)Un (x, r)hn (r)kdr
+
(9)
Now, A(x, D)Un (x, r)hn (r) Dom(p(D)) since en is a bounded subset of [0, ). By
Condition (A, p), then A(x, D)Un (x, r)hn (r) Dom(A(t, D)) Dom(f (t, D))
Dom(g (t, D)) for each t [0, T ], and by standard properties of evolution systems
(cf. [16, Theorem 5.1.2]),
Z x
k(I W,n (x, r))A(x, D)Un (x, r)hn (r)kdr
Zs x
=
k(W,n (x, x) W,n (x, r))A(x, D)Un (x, r)hn (r)kdr
Zs x
Z x
=
W,n (x, q)A(x, D)Un (x, r)hn (r)dq
dr
s
r q
Z x
Z x
(W,n (x, q)g (q, D)En )A(x, D)Un (x, r)hn (r)dq
=
dr
Zs x Z xr
kW,n (x, q)g (q, D)A(x, D)Un (x, r)hn (r)kdq dr.
s
r
Rt
Recall from the paragraph before Corollary 2.4 that W,n (t, s) = e s g (,D)d on
En H. By the assumption that g (t, ) for all (t, ) [0, T ] [0, ) and by
Condition (A, p), we then have
Z xZ x
kW,n (x, q)g (q, D)A(x, D)Un (x, r)hn (r)kdq dr
Zs x Zr x
Similarly,
k(I W,n (x, s))A(x, D)Un (x, s)n k T (1 + eT )kp(D)A(x, D)Un (x, s)n k,
and so by calculation (9), we have shown
kA(x, D)un (x) A(x, D)v,n (x)k
T (1 + eT )kp(D)A(x, D)Un (x, s)n k
Z x
+
T (1 + eT )kp(D)A(x, D)Un (x, r)hn (r)kdr.
s
(10)
266
MATTHEW A. FURY
(11)
(12)
(13)
n () is absolutely convergent,
such that
Z
1
1
1
(14)
z + z + 1 + dy K 1 + log |x t|
Z
)kp(D)Un (t, s)n k +
K 0
(15)
267
kn ()k =
+
dxdy
z z + 1 +
S
Z T Z
1
1
1
dy dx
C 0
+
z + 1 +
(16)
z
s
Z T
1
K
1 + log
C0
dx
|x t|
s
C 0
for a possibly different constant C 0 independent of , , and n. Therefore, from
(15) and (16), we have for = t + i S,
kn ()k kn ()k + kn ()k (K 0 + C 0 ),
(17)
ts
(18)
v
(s))
(s
+
i)
|(
(s
+
i),
)|
k
(s
+
i)kkk
=
n
,n
n
n
n
iD
e (n n ) n (s + i)
kk = kn (s + i)kkk C 0 kk, and so
Mn (s) = max |(n (s + i), )| C 0 kk.
(19)
Next, for R, using (17) and the fact that 0 < < 1, we have |(n (T + i), )|
kn (T + i)kkk (K 0 + C 0 )kk, and so
Mn (T ) = max |(n (T + i), )| M kk.
(20)
ts
T t
st
ts
T t
ts
= (C + T s M T s C 0 ) T s M T s C T s M T s
for a possibly different constant C independent of and n. Letting n , we
have found constants C and M , independent of , such that
T t
ts
ku(t) v (t)k C T s M T s
for 0 s t < T as desired.
268
MATTHEW A. FURY
4. Regularization for problem (3). In this section, we utilize the results from
Theorem 3.1 to prove the main result of the paper, that is regularization for problem
(3). Following [10, Definition 3.1], we have
Definition 4.1. A family {V (t, s) | > 0, 0 s t T } B(H) is called a
family of regularizing operators for the problem (3) if for each classical solution u(t)
of (3) with initial data Y and h C([s, T ] : Y ), and for any > 0, there exists
() > 0 such that
(i) () 0 as 0,
Rt
(ii) ku(t) (V() (t, s) + s V() (t, r)h(r)dr)k 0 as 0 for s t T
whenever k k .
Rt
We show that the evolution system V (t, s) = e s f (,D)d given in Proposition 2.1 provides a family of regularizing operators for the problem (3) where the
function f (t, ) is given by two different approximations each satisfying Condition
(A, p). The first approximation f : [0, T ] [0, ) R, motivated by work of
Miller [15], Lattes and Lions [12], Ames [2], and Ames and Hughes [3] will be given
by
f (t, ) =
k
X
aj (t)j k+1 .
(21)
j=1
Rt
s
f (,D)d
| > 0, 0 s t T }
= k
.
(k + 1)
(k + 1)
(k + 1)k+1
k+1 k
Setting = kB(k+1)kk+1 , we have that f (t, ) for small and so Proposition 2.1 (i) is satisfied. Also,
|f (t, )| r () := b() + k+1
(23)
269
for all (t, ) [0, T ] [0, ) and Dom(f (t, D)) = Dom(r (D)) = Dom(Dk+1 ) for
all t [0, T ]. It is easily shown that Proposition 2.1 (iii) holds (cf. [6, Example 4.1]).
Hence if h(t) is continuous from [s, T ] to Y = Dom(r (D)) in the graph norm k kY
associated with the operator r (D), then the hypotheses (i)(iv) of Proposition 2.1
hold so that (4) is well-posed.
It is easy to check that f satisfies Condition (A, p) with p() = k+1 . Also,
g (t, ) = k+1 0 for all (t, ) [0, T ] [0, ), so we may choose = 0.
T t
ts
Theorem 3.1 then yields the result ku(t)v (t)k C T s M T s for 0 s t < T ,
where u(t) and v (t) are classical solutions of (3) and (4) respectively.
Approximation (22): For the second approximation, we have for small ,
f (t, ) r () :=
B
b()
1 + k
(24)
Pk
where again B = j=1 bj . Setting = B
, we have that f (t, ) for small
and so Proposition 2.1 (i) is satisfied. Since r () is a bounded Borel function on
[0, ), the Spectral Theorem implies that r (D) = b(D)(I + Dk )1 is a bounded
everywhere-defined operator on H and so we may choose Y = Dom(r (D)) = H.
It is easy to check then that hypotheses (ii) and (iii) of Proposition 2.1 hold where
f (t, D) = A(t, D)(I + Dk )1 B(H) for each t [0, T ]. Therefore, if h
C([s, T ] : H), then Proposition 2.1 (i)(iv) hold so that (4) is well-posed.
Pk
It is not hard to show that f satisfies Condition (A, p) with p() = j=1 bj k+j
(cf. [6, Example 4.2]). Further, g (t, ) = A(t, ) + A(t, )(1 + k )1 0 for
all (t, ) [0, T ] [0, ), so we may choose = 0. Again, Theorem 3.1 yields the
ts
T t
result ku(t) v (t)k C T s M T s for 0 s t T .
Regularization: Now, for regularization, we assume u(t) is a classical solution of (3)
where Y and h C([s, T ] : Y ). Let > 0 be given and assume k k .
Also, let v (t) be a classical solution of (4) with f defined by either (21) or (22).
Rt
By Proposition 2.1, then v (t) = V (t, s) + s V (t, r)h(r)dr. Note, for x H and
0 s t T,
Z R
t
kV (t, s)xk2 =
|e s f (,)d |2 d(E()x, x)
0
Z
T 2
(e
)
d(E()x, x)
0
= (e T )2 kxk2
so that kV (t, s)k e T . Hence, for 0 s t T , we have
Z t
ku(t) (V (t, s) +
V (t, r)h(r)dr)k
s
Z t
ku(t) v (t)k + kv (t) (V (t, s) +
V (t, r)h(r)dr)k
s
(25)
270
MATTHEW A. FURY
ku(t) (V (t, s) +
V (t, r)h(r)dr)k
s
T t
B k+1 kk T
ts
C T s M T s + e k (k+1)k+1
t
k1 TT s
ts
2B k+1 k k T
=C
M T s +
k+1
(k + 1)
ln
0 as 0.
(26)
In the case of the second approximation (22), choose = 2BT /ln . Then 0
as 0 and
Z t
V (t, r)h(r)dr)k
ku(t) (V (t, s) +
s
T t
T s
ts
T s
+ e
=C(2BT /ln )
0
T t
T s
BT
ts
T s
(27)
as 0.
Finally, for the case that t = T , from inequalities (17) and (18), it is easily shown
by an argument similar to that in the last paragraph of the proof of Theorem 3.1
with t = T that ku(T ) v (T )k N for some constant N independent of . Then
by (25), in the case of either approximation (21) or (22), we have that 0 as
0 and
Z T
ku(T ) (V (T, s) +
V (T, r)h(r)dr)k
s
T
ku(T ) v (T )k + e
N +
0
as
(28)
0.
Combining (26), (27), and (28), we have in the case of either approximation, that
{V (t, s) | > 0, 0 s t T } is a family of regularizing operators for problem
(3).
5. Examples. The main application of our theory is to partial differential equations in the Hilbert space (H, k k) = (L2 (Rn ), k k2 ) with D = where
Pn
2
denotes the Laplacian defined by = i=1 x2 . The operator is a positive,
i
(t, x) [s, T ) Rn
(29)
Pk
in L2 (Rn ), where A(t, ) = j=1 aj (t)()j and aj C([0, T ] : R+ ) for 1 j
k, is generally ill-posed. By the results in Section 4, regularization for (29) stems
from the approximate well-posed problem (4) where f (t, D) = f (t, ) is defined
by two different functional calculi f :
271
Example 5.1. By the first approximation f (t, ) = A(t, ) k+1 , the approximate well-posed problem (4) becomes
x Rn .
Here, as in the proof of Theorem 4.2, we note that Y = Dom(r ()) = Dom(k+1 )
where r () = b() + k+1 as in (23), and Y is the Sobolev space W 2(k+1),2 (Rn )
consisting of functions L2 (Rn ) whose derivatives, in the sense of distributions
of order j 2(k + 1) are in L2 (Rn ) (cf. [16, Section 7.1]). Hence, in accordance
with Proposition 2.1 (iv), we require t h(t, ) to be continuous in the graph norm
k kY associated with the operator r () so that (30) is indeed well-posed.
For a simple example, consider problem (29) in H = L2 (R), where A(t, ) =
2
a(t)() with a C([0, T ] : R+ ), and h(t, x) is defined by h(t, x) = e(xt) . Thus,
the ill-posed problem (29) becomes the inhomogeneous backward heat equation
2
2
(t, x) [s, T ) R
(t, x) [s, T ) R
(31)
It is easy to check that e(xt) , and its derivatives with respect to x up to the
2
fourth order are each in L2 (R) since q(x)e(xt) L2 (R) for any polynomial q(x).
2
Using a dominated convergence argument, it may also be shown that t e(xt)
is continuous in the graph norm k kY associated with the operator r () =
B + 2 . Hence, (31) is well-posed.
Returning to the general case, as seen in Section 4, Theorem 4.2 implies that the
approximation
f (t, ) = A(t, ) k+1 of Example 5.1 yields an evolution system
Rt
V (t, s) = e s f (,)d , 0 s t T which serves as a family of regularizing
operators for the ill-posed problem (29).
A(t,)
Example 5.2. Using the second approximation f (t, ) = 1+
k , the approximate
well-posed problem (4) becomes
v(s, x) = (x),
x Rn .
b()
In this case, r () = 1+
k as in (24), and since r () is a bounded everywheredefined operator, we have Y = Dom(r ()) = L2 (Rn ). Thus, we require t
h(t, ) to be continuous in the ordinary L2 -norm k k2 so that (32) is indeed wellposed.
272
MATTHEW A. FURY
A(t,)
1+k
Rt