Samarskii A.A., Vabishchevich P.N. Numerical Methods for Solving Inverse Problems of Mathematical Physics

Подождите немного. Документ загружается.

306 Chapter 7 Evolutionary inverse problems

On the top and bottom sides, the following boundary conditions are considered:

∂u

∂t

(x, 0) = 0, 0 ≤ x ≤ l, (7.182)

u(x, T ) = u

(x), 0 ≤ x ≤ l. (7.183)

Among the inverse evolutionary problems for equation (7.180), we can identify the

Cauchy problem and the continuation problem for the solution of the boundary value

problem. The Cauchy problem is formulated as follows. Suppose that the boundary

condition at t = T is not speciﬁed, but we know the solution at t = 0, i.e., instead of

(7.183), the following condition is considered:

u(x, 0) = u

(x), 0 ≤ x ≤ l. (7.184)

In practical cases, of interest can be the problem in which it is required to protract the

solution of the direct problem (7.180)–(7.183) beyond the calculation-domain bound-

ary. For instance, we are interested in the solution u(x, t) in the domain

T +T

with

T > 0, i.e., here the solution is to be continued into a region adjacent to the top

portion of the boundary. After the solution of the problem inside the calculation do-

main

is found, the continuation problem reduces to a Cauchy problem similar to

problem (7.180)–(7.182), (7.184). Also of interest is the possibility of passing from

the Cauchy problem to a continuation problem. With this remark, in what follows we

will concentrate our attention on the Cauchy problem.

We will consider problem (7.180)–(7.182), (7.184) from a somewhat more general

standpoint as the Cauchy problem for the second-order differential-operator equation.

For functions given on the interval  = (0, 1) we deﬁne, in the traditional manner,

the Hilbert space H = L

(). On the set of functions vanishing at ∂ we deﬁne the

operator

Au =−

∂

∂x

, 0 < x < l. (7.185)

Among the key properties of this operator, note that in H we have:

∗

= A ≥ mE, m > 0. (7.186)

Equation (7.180), supplemented with condition (7.181) at the boundary, can be written

as the following differential-operator equation for u(t) ∈ H:

− Au = 0, 0 < t ≤ T. (7.187)

Initial conditions (7.182) and (7.184) yield:

u(0) = u

, (7.188)

(0) = 0. (7.189)

Section 7.4 Second-order evolution equation 307

In (7.187), the operator A is a self-adjoint, positively deﬁned operator. Problem

(7.187)–(7.189) is an ill-posed problem because, here, continuous dependence on in-

put data (initial conditions) is lacking. Conditional well-posedness takes place in the

class of solutions bounded in H.

7.4.2 Equivalent ﬁrst-order equation

In the construction of regularizing algorithms for the approximate solution of problem

(7.187)–(7.189), passage to a Cauchy problem for the ﬁrst-order evolution equation

may prove useful. In the latter case, we can follow the above-considered approaches

to the solution of ill-posed problems for evolution equations using perturbed initial

conditions and/or perturbed equation.

The simplest transformation related with the traditional introduction of the vector

of unknown quantities U ={u

, u

}, u

= u, u

= du/dt results in a system of ﬁrst-

order equations with a non-self-adjoint operator. This approach will be discussed in

more detail below.

In the case of problem (7.187)–(7.189), the self-adjointness and positive deﬁniteness

of A can be taken into account.

We perform the following change of variables:

v(t) =



u − A

−1/2



,w(t) =



u + A

−1/2



. (7.190)

Then, from (7.187) it readily follows that the new unknown quantities v(t) and w(t)

satisfy the following ﬁrst-order equations:

+ A

1/2

v = 0,

− A

1/2

w = 0. (7.191)

With regard to (7.188), (7.189) and for the introduced notation (7.190), for equations

(7.191) we pose the initial conditions

v(0) =

,w(0) =

. (7.192)

Thus, starting from the ill-posed problem (7.187)–(7.189), we arrive at a well-posed

problem in which it is required to determine v(t), and also at an ill-posed problem for

w(t). Hence, regularizing algorithms for problem (7.187)–(7.189) can be constructed

based on regularization of the split system (7.191). Below, in the consideration of

the generalized inverse method for problem (7.187)–(7.189), we will outline some

possibilities available along this direction. In practical realizations, it seems reasonable

to use perturbations related with the calculation of the square root of A.

We deﬁne the vector U ={u

, u

} and the space H

as the direct sum of spaces H:

= H ⊕ H. The addition in H

is to be performed coordinatewise, and the scalar

product there is deﬁned as follows:

(U, V ) = (u

) + (u

308 Chapter 7 Evolutionary inverse problems

Suppose that u

= v, u

= w; then, system (7.191) assumes the form

− LU = 0, (7.193)

where

L =

−A

1/2

0 A

1/2

. (7.194)

Equation (7.194) is supplemented with the boundary condition (see (7.192))

U (0) = U

, U





. (7.195)

Problem (7.193), (7.195) belongs to the above-considered class of ill-posed prob-

lems for the ﬁrst-order equation. The speciﬁc feature of the problem is due to fact that

the operator L (see (7.194)) may change its sign.

Consider some possibilities in the passage to an equivalent system of ﬁrst-order

equations under more general conditions, for instance, in the case in which the oper-

ator A is a non-self-adjoint operator. Suppose that the components of U ={u

, u

}

from H

are deﬁned as

= u, u

. (7.196)

Then, the initial problem (7.187)–(7.189) can be written as (compare with (7.193)–

(7.195)) equation (7.193), supplemented now with the initial condition

U (0) = U

, U

={u

, 0}. (7.197)

For the operator L we obtain the representation

L =

0 E

A 0

. (7.198)

Thus, here again from the Cauchy problem for the evolution second-order equation we

have passed to the Cauchy problem for the evolution ﬁrst-order equation.

7.4.3 Perturbed initial conditions

Let us brieﬂy outline the available possibilities in using methods based on non-local

perturbation of initial conditions as applied to the approximate solution of the ill-posed

Cauchy problem for the second-order evolution equation. Generally speaking, in the

case of general problems, two initial conditions are to be perturbed. Suppose that

we pose ﬁrst an ill-posed problem in which it is required to determine the function

u = u(t) ∈ H from equation (7.187) supplemented with initial conditions (7.188) and

(7.189).

Section 7.4 Second-order evolution equation 309

The approximate solution u

(t) is deﬁned as the solution of the following non-local

problem:

− Au

= 0, 0 < t ≤ T, (7.199)

(0) + αu

(T ) = u

, (7.200)

(0) = 0. (7.201)

A speciﬁc feature in the non-local problem (7.199)–(7.201) consists in the fact that

here only one initial condition (see (7.200)) is to be perturbed. The stability estimate

is given by the following statement.

Theorem 7.20 For the solution of the non-local problem (7.199)–(7.201) there holds

the estimate

u

(t)≤

u

. (7.202)

Proof. The solution of problem (7.199)–(7.201) can be written in the traditional oper-

ator form

(t) = R(t,α)u

, (7.203)

where

R(t,α) = ch (A

1/2

t)(E + α ch (A

1/2

T ))

−1

. (7.204)

We assume that the spectrum of A is discrete, consisting of eigenvalues 0 <λ

≤

≤···, and the related system of eigenfunctions {w

}, w

∈ D(A), k = 1, 2,...is

an orthonormal complete system in H. In such conditions, for the solution of problem

(7.199)–(7.201) we have the representation

(t) =

∞



k=1

) ch (λ

1/2

t)(1 + α ch (λ

1/2

T ))

−1

From here, the stability estimate (7.203) follows.

Using the perturbed initial condition, one can establish regularizing properties of the

algorithm perfectly analogous to the case of the ﬁrst-order equation (see Theorem 7.4).

Suppose that, instead of the exact initial condition u

, an approximate solution u

given. As usually, we assume that there exists a stability estimate of form

u

− u

≤δ. (7.205)

The approximate solution for inaccurate initial conditions is to be found from equation

(7.199), condition (7.201), and from the non-local condition

(0) + αu

(T ) = u

. (7.206)

310 Chapter 7 Evolutionary inverse problems

Theorem 7.21 With estimate (7.205) fulﬁlled, the solution of problem (7.199),

(7.201), (7.206) converges to the solution of problem (7.187)–(7.189), bounded in H,

provided that δ → 0, α(δ) → 0, δ/α → 0.

Consider the relation between the non-local problem (7.199)–(7.201) and the opti-

mal control problem. In the case of interest, both the solution of the optimal control

problem and the solution of the non-local problem are obtained based on the variable

separation method.

Let v ∈ H be the sought control and u

(t;v) be deﬁned as the solution of the

problem

− Au

= 0, 0 < t ≤ T, (7.207)

(T ;v) = v, (7.208)

(0) = 0. (7.209)

We consider the quality functional in the following simplest form:

(v) =u

(0;v) − u



+ αv

. (7.210)

For the optimal control w,wehave:

(w) = min

v∈H

(v). (7.211)

The solution of the optimal control problem (7.207)–(7.211) can be written in the

form (7.203) provided that

R(t,α) = ch (A

1/2

t)(E + α ch

1/2

T ))

−1

. (7.212)

Through comparison of (7.204) with (7.212), the following statement can be estab-

lished.

Theorem 7.22 The solution of the optimal control problem (7.207)–(7.211) coincides

with the solution of the non-local problem for equation (7.199) on the double interval

(0, 2T ) with initial condition (7.201) and non-local

(0) +

2 + α

(2T ) =

2 + α

. (7.213)

Proof. To derive (7.213) we use for the solution the representation (7.212) and formula

2ch(x ) = ch (2x) + 1.

Some other non-local conditions can be obtained by writing the Cauchy problem

(7.187)–(7.189) as a Cauchy problem for the ﬁrst-order evolution equation. An exam-

ple here is problem (7.193)–(7.195).

Section 7.4 Second-order evolution equation 311

We determine the approximate solution of problem (7.193)–(7.195) as the solution

of the non-local problem

− LU

= 0, (7.214)

(0) + αU

(T ) = U

. (7.215)

To formulate the related non-local problem for the second-order equation, in prob-

lem (7.214), (7.215) we perform the reverse transition. By analogy with (7.190), we

represent the solution of (7.214) as

(t) =



− A

−1/2



(t) =



+ A

−1/2



(7.216)

From (7.216) it follows that

(t) = u

(t) + u

(t),

(t) = A

−1/2

(t) − u

(t)).

(7.217)

Then, with (7.195) the non-local condition (7.215) yields:

(0) + αu

(T ) =

(0) + αu

(T ) =

(7.218)

With (7.216), (7.217) taken into account, from (7.218) we obtain:

(t) + αu

(T ) = u

(0) +

(T ) = 0.

(7.219)

We arrive at two non-local conditions for the solution u

(t) of equation (7.199). Here

(compare with (7.200) and (7.201)), two, instead of one, initial conditions are to be

perturbed.

7.4.4 Perturbed equation

Consider the possibilities available in the construction of regularized solution algo-

rithms suitable for inverse problems for second-order equations based on the perturbed

initial equation. Here, we can use the above-considered versions of the generalized in-

verse method for ﬁrst-order evolution equations.

312 Chapter 7 Evolutionary inverse problems

To approximately solve problem (7.187)–(7.189), we use the generalized inverse

method in its version analogous to the basic version of this method for ﬁrst-order

equations. The approximate solution u

(t) is to be found from the equation

− Au

+ αA

= 0, 0 < t ≤ T. (7.220)

The initial conditions are given with some inaccuracy. Suppose that

(0) = u

, (7.221)

(0) = 0. (7.222)

We assume that for the initial-condition inaccuracy the estimate (7.205) holds.

The solution of (7.220)–(7.222) can be written in the operator form

(t) = R(t,α)u

, (7.223)

where

R(t,α) = ch ((A − αA

)

1/2

t).

Using the variable separation method, we can prove convergence of the approximate

solution to the exact solution.

Theorem 7.23 Let for the initial-condition inaccuracy the estimate (7.205) hold.

Then, in the case of δ → 0, α(δ) → 0, δ exp (T/(4α)

1/2

) → 0 the approximate

solution u

(t), determined as the solution of problem (7.220)–(7.222), converges to

the exact solution u(t) of problem (7.187)–(7.189), bounded in H, and the following

stability estimate with respect to initial data holds:

u

(t)≤ch



√



u

(0). (7.224)

Proof. To prove the statement and, in particular, to show the validity of esti-

mate (7.224), we can follow the procedure previously used to prove analogous results

for the ﬁrst-order equation (see Theorem 7.12, for instance).

The version of the generalized inverse method for equation (7.187) analogous to

pseudo-parabolic perturbation of the ﬁrst-order equation consists in the determination

of the approximate solution from the equation

− Au

+ αA

= 0, 0 < t ≤ T, (7.225)

supplemented with initial conditions (7.221), (7.222).

Section 7.4 Second-order evolution equation 313

The approximate solution converges to the exact solution provided that the value of

the regularization parameter is matched with the input-data inaccuracy, and for the so-

lution of problem (7.221), (7.222), (7.225) there holds the following stability estimate

with respect to initial data:

u

(t)≤ch



√



u

(0). (7.226)

Variants of the generalized inverse method for the approximate solution of the ill-

posed Cauchy problem (7.187)–(7.189) can be constructed based on the passage to an

equivalent Cauchy problem for the system of ﬁrst-order equations. Instead of (7.187)–

(7.189), consider the problem (7.193)–(7.195), in which

(t) = v(t) =



u − A

−1/2



(t) = w(t) =



u + A

−1/2



(7.227)

Taking the fact into account that the operator L is a self-adjoint operator that can

change its sign, we ﬁnd the approximate solution U

(t) of problem (7.193), (7.195)

from the equation

− LU

+ αL

= 0 (7.228)

and the initial conditions

(0) = U

. (7.229)

Based on the previously obtained results (see Theorem 7.11), derive the following

estimate of stability of the approximate solution U

(t) with respect to initial data:

U

(t)≤exp



4α



U

(0). (7.230)

First, we derive an equation from which the approximate solution u

(t) that corre-

sponds to system (7.228) can be found. In view of (7.194), we have

A 0

0 A

With regard to (7.228), for v

(t), w

(t) (with U

={v

}) we obtain the following

system of equations:

+ A

1/2

+ αAv

= 0,

− A

1/2

+ αAw

= 0.

(7.231)

By analogy with the exact solution u(t) (see (7.227)), we deﬁne the approximate

solution u

(t) by the following equation:

(t) = v

(t) + w

(t). (7.232)

314 Chapter 7 Evolutionary inverse problems

From (7.231) and (7.232), it follows immediately that the function u

(t) satisﬁes the

equation

− Au

+ 2αA

+ α

= 0, 0 < t ≤ T. (7.233)

It follows from (7.232) that the solution of equation (7.233), supplemented with ap-

propriate initial conditions, satisﬁes

u



≤ (v

+w

)

≤ 2(v



+w



) = 2U



≤ 2exp



2α



U

(0)

≤ exp



2α



u

(0)

Of course, the same estimate can be obtained starting from the estimates for v

(t) and

(t) derived from equation (7.231).

Equation (7.233) was obtained by perturbing two terms. The regularization per-

formed by perturbing one of the terms (α

) was considered previously (see equa-

tion (7.220)). Of interest here is to consider, in its pure form, the regularization per-

formed at the expense of the second term. For this reason, we will seek the approxi-

mate solution from the equation (see (7.233))

− Au

+ αA

= 0, 0 < t ≤ T, (7.234)

supplemented with the initial conditions (7.221), (7.222). Such regularization, applied

to well-posed problems for evolution equations, is called “parabolic” regularization.

Similarly to Theorem 7.23, we can prove an analogous statement about regularizing

properties of the generalized inverse method in its “parabolic”-regularization version.

Here, we give only the stability estimate with respect to initial data:

u

(t)≤exp





u

(0).

We have restricted ourselves to the application of the generalized inverse method in its

standard version (7.228), (7.229) applied to problem (7.193)–(7.195). Some additional

possibilities are given by pseudo-parabolic perturbation of equation (7.193).

7.4.5 Regularized difference schemes

Let us turn now to constructing regularized difference schemes for the approximate

solution of the ill-posed Cauchy problem for the second-order evolution equation.

On the interval

 = [0, l], we introduce a uniform grid with a grid size h:

¯ω ={x | x = x

= ih, i = 0, 1,...,N, Nh = l}.

Here, ω is the set of internal nodes and ∂ω is the set of boundary nodes.

Section 7.4 Second-order evolution equation 315

At the internal nodes, we approximate the differential operator (7.185), accurate to

second order, with the difference operator

y =−y

¯xx

, x ∈ ω. (7.235)

In the mesh Hilbert space H, we introduce the norm by the relation y=(y, y)

1/2

in which

(y,w) =



x∈ω

y(x)w(x)h.

On the set of functions vanishing on ∂ω, for the self-adjoint operator  there holds the

estimate

 = 

∗

≥ λ

E, (7.236)

in which

sin

πh

≥

On the approximation over the space, to problem (7.187)–(7.189) we put into cor-

respondence the problem

− y = 0, x ∈ ω, t > 0, (7.237)

y(x, 0) = u

(x), x ∈ ω, (7.238)

(x, 0) = 0, x ∈ ω. (7.239)

Regularized difference schemes for problem (7.237)–(7.239) can be constructed

based on the regularization principle for difference schemes. As a generic (initial)

difference scheme, we adopt the explicit symmetric difference scheme

n+1

− 2y

+ y

n−1

− y

= 0,

x ∈ ω, n = 1, 2,...,N

− 1

(7.240)

with some initial conditions.

With regard to (7.238), we have:

= u

(x), x ∈ ω.

The solution at the time t = τ can be approximated, with conditions (7.238), (7.239)

on the solutions of (7.237). Taking into account the relation

= y

+ τ

(0) +

(0) + O(τ

to approximate (7.239), we use the difference relation

− y

y

= 0.