UDC 517.948.32
MATHEMATICS
Submitted 1968-01-01 | RussiaRxiv: ru-196801.87142 | Translated from Russian

Full Text

UDC 517.948.32

MATHEMATICS

V. N. STRAKHOV

ON THE NUMERICAL SOLUTION OF ILL-POSED PROBLEMS REPRESENTED BY INTEGRAL EQUATIONS OF CONVOLUTION TYPE

(Presented by Academician A. N. Tikhonov on 13 III 1967)

Many problems of exploration geophysics \(\left({}^{1,2}\right.\) and others) lead to linear integral equations of convolution type

\[ f(x)=\frac{1}{\sqrt{2\pi}}\int_{-\infty}^{+\infty}\varphi(\xi)K(x-\xi)\,d\xi,\qquad -\infty \leq x \leq +\infty, \tag{1} \]

whose kernels satisfy the conditions:

\[ \begin{aligned} &1)\quad |k(t)|\leq 1,\qquad -\infty \leq t \leq +\infty;\\ &2)\quad k(t)\ne 0\ \text{for almost all } t,\ -\infty<t<+\infty. \end{aligned} \tag{2} \]

Here

\[ k(t)=\frac{1}{\sqrt{2\pi}}\int_{-\infty}^{+\infty}K(x)e^{-itx}\,dx \tag{3} \]

is the spectrum (Fourier transform) of the kernel \(K(x)\).

We write equation (1) in operator form:

\[ f=A\varphi;\qquad \|A\|=\sup_t |k(t)|=1, \tag{4} \]

where \(A\) is a linear integral operator acting in the space \(L^2\), from the set \(P\), singled out by condition 2); \(f\) and \(\varphi\) are functions from the space \(L^2\). Obviously, there exists an inverse operator \(A^{-1}\) to \(A\), but

\[ \|A^{-1}\|=\sup_t |1/k(t)|=+\infty, \tag{5} \]

so that the solutions of equation (1) possess instability: small (in the norm in \(L^2\)) variations \(\delta f\) may correspond to large variations \(\delta\varphi\).

A. N. Tikhonov \(\left({}^{3,4}\right)\) proposed a method for the approximate solution of problems of the type (4)—(5)—the regularization method. It consists in introducing, instead of equation (4), a linear operator \(T_\alpha\) in the space \(L^2\), depending on a parameter \(\alpha\), \(0<\alpha<\alpha_0\), and satisfying the conditions:

\[ \begin{aligned} &1)\quad \|T_\alpha\|<+\infty,\qquad 0<\alpha<\alpha_0;\\ &2)\quad \|\varphi-T_\alpha A\varphi\|\xrightarrow[\alpha\to 0]{}0\quad (\text{for any } \varphi\in L^2). \end{aligned} \tag{6} \]

\[ \tag{7} \]

The function \(T_\alpha f\), for some \(\alpha=\alpha^*\), is taken as an approximate solution of equation (4). V. K. Ivanov \(\left({}^{5}\right)\) introduced the notion of uniform regularization. Namely, the operator \(T_\alpha\) uniformly regularizes equation (4) on some set \(m\subseteq L^2\), if, as \(\alpha\to 0\), \(\|\varphi-T_\alpha A\varphi\|\to 0\) uniformly for all \(\varphi\in m\).

If an equation \(f_\delta=A\varphi\) is given, where \(f_\delta=f+\delta f\), \(\|\delta f\|\leq \delta\), and moreover \(f\in M\), \(\delta f\in M\), then \(f_\delta\in M\), and a solution of the equation does not exist for any \(\delta>0\). However, an approximate solution can always be constructed, pro-

than

\[ \|\varphi-T_{\alpha}f_{\delta}\|\leq \|\varphi-T_{\alpha}A\varphi\|+\|T_{\alpha}\|\delta, \tag{8} \]

and by an appropriate choice of \(\alpha\) one can achieve sufficiently good results.

Definition 1. The value of the parameter \(\alpha\) determined from the condition

\[ \sup_{\varphi\in m}\|\varphi-T_{\alpha}A\varphi\|+\|T_{\alpha}\|\delta=\min \tag{9} \]

is called optimal on the set \(m\) of uniform regularization for a given \(\delta>0\).

Accordingly, \(T_{\alpha}f_{\delta}\), where \(\alpha\) is chosen according to (9), will be an optimal (on \(m\) for the given \(\delta\)) solution.

Definition 2. A linear operator \(L\), \(\|L\|<+\infty\), acting in the space \(L^2\), will be called computationally finite-dimensional if, in order to find the function \(Lf\), \(f(x)\in L^2\), at one value of the argument \(x\), a finite number of values \(f(x)\) is used, on which a finite number of arithmetic operations is performed. Operators \(L\) of the opposite kind will be called computationally infinite-dimensional.

If \(T_{\alpha}\) is a regularizing operator, but computationally infinite-dimensional, then constructing the function \(T_{\alpha}f\) even for a single \(x\) is practically impossible. Therefore one proceeds as follows: the operator \(T_{\alpha}\) is approximately approximated by some computationally finite-dimensional operator \(\widetilde T_{\alpha}\), and the approximate solution of the problem is taken to be not \(T_{\alpha}f\), but \(\widetilde T_{\alpha}f\). However, if the method for constructing the operator \(\widetilde T_{\alpha}\) from the operator \(T_{\alpha}\) is not regular, in a certain way connected with the value of \(\alpha\) used, then convergence may be lost: although \(T_{\alpha}A\varphi\to\varphi\) as \(\alpha\to0\), it may be that \(\widetilde T_{\alpha}A\varphi\to\widetilde\varphi\), \(\|\varphi-\widetilde\varphi\|>0\). Hence it follows that the procedure of passing from the computationally infinite-dimensional operator \(T_{\alpha}\) to the computationally finite-dimensional \(\widetilde T_{\alpha}\) must also be regularized. Such an operation is most simply carried out by introducing a second parameter \(\beta=n\) (\(n\) natural) and putting \(\widetilde T_{\alpha}=T_{\alpha,n}\), where, as \(n\to\infty\), \(T_{\alpha,n}f\to T_{\alpha}f\) for any \(f\in L^2\).

Definition 3. A linear operator \(T_{\alpha,n}\), \(0\leq \alpha\leq \alpha_0\), \(n_0\leq n\leq +\infty\), is called regularizing equation (4) on the space \(L^2\) if: 1) for all \(n\) and \(\alpha\), except the case when simultaneously \(n=+\infty\) and \(\alpha=0\),

\[ \|T_{\alpha,n}\|<+\infty; \tag{10} \]

2) if at least for one sequence of values \(\alpha\) and \(n\)

\[ \|\varphi-T_{\alpha,n}A\varphi\|\xrightarrow[\alpha\to0,\ n\to\infty]{}0 \quad \text{for any } \varphi\in L^2. \tag{11} \]

Definition 4. The operator \(T_{\alpha,n}\) is called uniformly regularizing equation (4) on the set \(m\in L^2\) if, for all \(\varphi\in m\) and at least for one sequence \((\alpha,n)\), \(\|\varphi-T_{\alpha,n}A\varphi\|\to0\) uniformly.

Definition 5. The values of the parameters \(\alpha\) and \(n\), determined from the condition

\[ \sup_{\varphi\in m}\|\varphi-T_{\alpha,n}A\varphi\|+\|T_{\alpha,n}\|\delta=\min, \tag{12} \]

are called optimal on the set \(m\) of uniform regularization for a given \(\delta>0\).

Theorem 1. In order that the two-parameter operator \(T_{\alpha,n}\) be regularizing, it is sufficient that \(T_{\alpha,\infty}=T_{\alpha}\) be a one-parameter regularizing operator and

\[ \|T_{\alpha}-T_{\alpha,n}\|\xrightarrow[n\to\infty]{}0 \tag{13} \]

uniformly on any interval \(0<\alpha_0^*\leq \alpha\leq \alpha_0\).

Proof follows from the inequality

\[ \|\varphi-T_{\alpha,n}A\varphi\|\leq \|\varphi-T_\alpha A\varphi\|+\|T_\alpha-T_{\alpha,n}\|\,\|A\varphi\|. \tag{14} \]

Let us proceed to the construction of methods for the approximate solution of integral equations of convolution type (1)—(4) by means of computationally finitely supported two-parameter regularizing operators. We shall seek the latter in the class of operators

\[ S_{\alpha,n}=\sum_{-n}^{+n} c_k E_\alpha^k, \tag{15} \]

where \(E_\alpha\) is the shift operator, \(E_\alpha^k\{f(x)\}=f(x+k\alpha)\), and the \(c_k\) are numerical coefficients depending on \(\alpha\) and \(n\). The approximate solution \(\varphi_{\alpha,n}(x)\) of equation (1) on the basis of the operator \(S_{\alpha,n}\) is determined by the expression

\[ \varphi_{\alpha,n}(x)=\sum_{-n}^{+n} c_k f(x+k\alpha). \tag{16} \]

Clearly, under the conditions for the existence of a solution \(\varphi(x)\in L^2\) of equation (1),

\[ \|\varphi-\varphi_{\alpha,n}\|_{L_2} = \left( \int_{-\infty}^{+\infty} |F(t)|^2 \left| \frac{1}{k(t)}-\sum_{-n}^{+n} c_k e^{-ik\alpha t} \right|^2 \,dt \right)^{1/2}, \tag{17} \]

where \(F(t)\) is the Fourier transform of the function \(f(x)\).

A natural method is the choice of coefficients from the conditions of expanding the function \(1/k(t)\) in a Fourier series on the interval \(|t|\leq \pi/\alpha\):

\[ c_k=c_k(\alpha)=\frac{\alpha}{2\pi} \int_{-\pi/\alpha}^{+\pi/\alpha} \frac{e^{ik\alpha t}}{k(t)}\,dt, \qquad k=0,\pm 1,\pm 2,\ldots \tag{18} \]

Consider the subclass of integral equations (1)—(4) satisfying the additional conditions, fulfilled for all \(\alpha>0\):

\[ \mu(\alpha)=\sup_{|t|\leq \pi/\alpha}\left|\frac{1}{k(t)}\right|<+\infty; \tag{19} \]

\[ |1-k(t)/k_{\mathrm{per}}^{(\alpha)}(t)|\leq C,\qquad |t|>\pi/\alpha, \tag{20} \]

where \(C\) is an absolute constant, and \(k_{\mathrm{per}}^{(\alpha)}(t)\) is the \(2\pi/\alpha\)-periodic repetition of the values of the function \(k(t)\) on the interval \(|t|\leq \pi/\alpha\).

Theorem 2. In order that, under conditions (19)—(20), the operator \(S_{\alpha,n}\) (15) with coefficients (18) be regularizing, it is sufficient that the expansion

\[ \frac{1}{k_{\mathrm{per}}^{(\alpha)}(t)} = \sum_{-\infty}^{+\infty} c_k(\alpha)e^{-ik\alpha t}, \qquad -\infty\leq t\leq +\infty, \tag{21} \]

converge uniformly for all \(\alpha>0\).

Proof. The operator \(S_{\alpha,n}\) is bounded for arbitrary \(\alpha>0\), \(n<+\infty\):

\[ \|S_{\alpha,n}\|=\sup_t \left| \sum_{-n}^{+n} c_k e^{-ik\alpha t} \right| <+\infty. \tag{22} \]

The operator \(S_\alpha=S_{\alpha,\infty}\) is regularizing. Indeed,

\[ \|S_\alpha\| = \sup_{\|f\|\leq 1} \left( \int_{-\infty}^{+\infty} |F(t)|^2 \left| \frac{1}{k_{\mathrm{per}}^{(\alpha)}(t)} \right|^2 \,dt \right)^{1/2} = \sup_{|t|\leq \pi/2} \left| \frac{1}{k(t)} \right| = \mu(\alpha)<+\infty \tag{23} \]

and \((\Phi(t)\) is the Fourier transform of the function \(\varphi(x))\):

\[ \|\varphi-S_\alpha A\varphi\| = \left( \int_{-\infty}^{+\infty} |F(t)|^2 \left| \frac{1}{k(t)}-\frac{1}{k_{\mathrm{per}}^{(\alpha)}(t)} \right|^2 \,dt \right)^{1/2} \leq \]

\[ \leq C\left(\int_{-\infty}^{-\pi/\alpha} |\Phi(t)|^2\,dt+\int_{\pi/\alpha}^{+\infty} |\Phi(t)|^2\,dt\right)^{1/2}, \tag{24} \]

and since

\[ \int_{-\infty}^{+\infty} |\Phi(t)|^2\,dt<+\infty, \]

as \(\alpha\to 0\) the right-hand side in (24) tends to zero. Since, in addition,

\[ \begin{aligned} \|S_\alpha-S_{\alpha,n}\| &=\sup_{\|f\|\leq 1} \left( \int_{-\infty}^{+\infty} |F(t)|^2 \left| \frac{1}{k_{\mathrm{per}}^{(\alpha)}(t)} -\sum_{-n}^{+n} c_k e^{-ikat} \right|^2 dt \right)^{1/2} \\ &= \sup_{|t|\leq \pi/\alpha} \left| \frac{1}{k_{\mathrm{per}}^{(\alpha)}(t)} -\sum_{-n}^{+n} c_k e^{-ikat} \right|, \end{aligned} \tag{25} \]

and, by virtue of the uniform convergence (21), tends to zero as \(n\to\infty\), the validity of Theorem 2 follows from Theorem 1.

Let \(\varphi_\sigma(x)\in L^2\) be a collection of functions whose spectra \(\Phi_\sigma(t)\) satisfy the boundedness condition:

\[ \Phi_\sigma(t)\equiv 0 \quad \text{for } |t|>\sigma. \tag{26} \]

The set \(m\) of functions \(\varphi\) on which the two-parameter regularizing operator \(S_{\alpha,n}\) admits uniform regularization can be specified as the collection of functions with bounded spectrum (26), with \(\|\varphi_\sigma\|\leq N\).

In this case, for \(\varphi=\varphi_\sigma\in m\), we have the estimate

\[ \|\varphi_\sigma-S_{\alpha,n}A\varphi_\sigma\| \leq N\sup_{|t|\leq \sigma} \left| \frac{1}{k_{\mathrm{per}}^{(\alpha)}(t)} -\sum_{-n}^{+n} c_k e^{-ikat} \right| + \delta\sup_{|t|\leq \sigma} \left| \sum_{-n}^{+n} c_k e^{-ikat} \right|. \tag{27} \]

The operator \(S_{\alpha,n}\) optimal on the set \(\varphi_\sigma\in m\), for a given \(\delta\), is determined by the condition

\[ \frac{\delta}{N}\sup_{|t|\leq \sigma} \left| \sum_{-n}^{+n} c_k e^{-ikat} \right| + \sup_{|t|\leq \sigma} \left| \frac{1}{k_{\mathrm{per}}^{(\alpha)}(t)} -\sum_{-n}^{+n} c_k e^{-ikat} \right| = \min. \tag{28} \]

Up to now we have considered regularization on the whole space \(L^2\). However, regularization may also be considered on certain sets \(E\) of functions from \(L^2\) admitting a stronger metrization. In these cases, the operator regularizing on the set \(E\) is defined analogously, with the norm from \(L^2\) in (7) and (11) replaced by the stronger norm of this set.

Theorem 3. Let \(L^2_{(k+1)}\) be the set of functions differentiable \(k+1\) times, with derivatives belonging to \(L^2\). Under the condition \(\varphi(x)\in L^2_{(k+1)}\) and the fulfillment of the restrictions (19)—(21), the operator \(S_{\alpha,n}\) is regularizing in the sense of uniform convergence

\[ \max_x \left| d^r\varphi(x)/dx^r - d^r\varphi_{\alpha,n}(x)/dx^r \right| \longrightarrow 0,\quad r\leq k, \tag{29} \]

as \(\alpha\to 0,\ n\to\infty\), and convergence

\[ \left\| d^{k+1}\varphi(x)/dx^{k+1} - d^{k+1}\varphi_{\alpha,n}(x)/dx^{k+1} \right\|\to 0 \]

in the space \(L^2\).

Schmidt Institute of Physics of the Earth,
Academy of Sciences of the USSR

Received
16 II 1967

CITED LITERATURE

  1. M. M. Lavrent’ev, On Some Ill-Posed Problems of Mathematical Physics, Publishing House of the Siberian Branch of the Academy of Sciences of the USSR, 1962.
  2. M. M. Lavrent’ev, V. G. Vasil’ev, Sibirsk. Mat. Zhurn., 7, No. 3, 559 (1966).
  3. A. N. Tikhonov, DAN, 151, No. 3, 501 (1963).
  4. A. N. Tikhonov, DAN, 153, No. 1, 49 (1963).
  5. V. K. Ivanov, Sibirsk. Mat. Zhurn., 7, No. 3, 546 (1955).
  6. V. N. Strakhov, DAN, 153, No. 3, 533 (1963).

Submission history

UDC 517.948.32