Approximation of unbounded operators by bounded operators and related extremal problems

DOI10.1070/RM1996V051N06ABEH003001MaRDI QIDQ4375223zbMATH OpenOpenAlexFDO

Publication date 1 February 1998

Published in Russian Mathematical Surveys (Search for Journal in Brave)

Full work available at URL https://doi.org/10.1070/rm1996v051n06abeh003001

Research exposition (monographs, survey articles) pertaining to approximations and expansions (41-02) Best approximation, Chebyshev systems (41A50)

Review text

The unit sphere in \(L^p\oplus L^q\) \dots{} In this expository article, dedicated to the memory of Sergei Borisovich Stechkin, the author describes the current state of affairs with regards to the question of approximation of unbounded linear operators by bounded linear operators on a class of elements of a Banach space starting with the formulation of the problem as done by Stechkin in 1965. Notations. Let \(X,Y\) be Banach spaces, \(A: X\to Y\) an operator with domain \(D(A)\subset X, Q\) a class of elements in \(D(A)\). Let furthermore \({\mathbb L}(N)={\mathbb L}(N;X,Y)\) the set of bounded linear operators from \(X\) to \(Y\) with a norm \(\|T\|_{X\to Y}\) that does not exceed a number \(N\geq 0\). The quantity \[ U(T)=\sup\{\|Ax-Tx\|_Y \bigm|x\in Q\} \tag{0.1} \] is the deviation of the operator \(T\in{\mathbb L}(N)\) from \(A\) on the class \(Q\) and \[ E(N)=\inf\{U(T) \bigm|T\in {\mathbb L}(N)\} \tag{0.2} \] is the best approximation of \(A\) by the set of bounded operators \(\mathbb L\) on \(Q\). Stechkin's problem consists of estimating \(E(N)\), and studying the exsistence and uniqueness of an extremal operator. The author states in generality the various areas in which results have been obtained but concentrates on presenting Stechkin's results. The results in this paper have been classified in seven areas. \S 1. Bounds for the degree of best approximation of an operator. \S 2. The connection between Stechkin's problem and the ill-posed problem of recovering the values of an unbounded operator. \S 3. Best approximation of functionals. \S 4. Best approximation of differential operators and inequalities between norms of derivatives of functions of one variable on the line and a half-line. \S 5. Approximation of differential operators on classes of functions of several variables. \S 6. Approximation of invariant operators. \S 7. Approximation of differentiable functions by smoother functions. In each section a few main results shall be mentioned including the necessary definitions but no claim is made of a comprehensive overview. \S 1. The two functions \[ \begin{aligned}\Omega(\delta) &=\sup\{\|Ax_1-Ax_2\|_Y \bigm|x_1,x_2\in Q, \|x_1-x_2\|_X \leq\delta\}\\ \omega(\delta) &= \sup\{\|Ax\|_Y \bigm|x\in Q, \|x\|_X\leq\delta\}, \end{aligned} \] of a real variable \(\delta\in[0,\infty)\) are called modulus of continuity of \(A\) on \(Q\). If \(A\) is linear and \(Q\) is centrally symmetric and convex then \(\Omega(2\delta)=2\omega(\delta)\). In addition the notation \[ \begin{aligned} \Delta(N)=\Delta(N,Q)&=\sup\{\omega(\delta)-N\delta \bigm|\delta\geq 0\}\\ &=\sup\{\|Ax\|_Y-N\|x\|_X \bigm|x\in Q\},\\ \ell(\delta)&=\inf\{E(N)+N\delta \bigm|N>0\}. \end{aligned} \] Theorem. If \(A\) is a homogeneous (in particular linear) operator, and \(Q\) is a centrally symmetric convex subset of the domain \(A\), then the inequalities \[ E(N)\geq \Delta(N),\quad N\geq 0,\quad \omega(\delta)\leq \ell(\delta),\quad \delta\geq 0. \] hold. If in addition there exist an element \(x\in Q\) and a bounded linear operator \(T\) with the properties \[ \|Ax\|=U(T)+\|T\|\|x\|\tag{1} \] then the following equalities hold: \[ \omega(\|x\|)=\|x\|, \quad E(\|T\|)=U(T)=\|Ax\|-\|T\|\|x\| \] and hence the operator \(T\) is extremal for \(N=\|T\|\) and the element \(x\) is extremal with respect to \(\omega(\delta)\) for \(\delta=\|x\|\). The conditions for which the inequalities become equalities is still unsolved. \S 2. For \(\delta\geq 0\) and operators \(A,T: X\to Y\), a set \(Q\subset D(A)\) and a set of operators \({\mathbb R}: X\to Y\) define the two quantities \[ U_\delta(T)=U_\delta(T,A,Q)=\sup\{\|Ax-T_\eta\|_Y \bigm|x\in Q, \eta\in X, \|x-\eta\|_X\leq \delta\} \tag{2.1} \] and \[ E_\delta({\mathbb R})= E({\mathbb R};A,Q)= \inf \{U_\delta(T) \bigm|T\in {\mathbb R}\} \] \(E_\delta({\mathbb R})\) can be thought as the least error of recovery of the operator \(A\) via the set of maps \({\mathbb R}\) (method of recovery) on the elements of a class \(Q\) defined with an error of \(\delta\). As for the set \({\mathbb R}\) frequent canditates are the set of all maps \(X\to Y\), the linear maps, the bounded maps, the linear maps with norm not exceeding a number \(N\). These are denoted by \(\mathbb O,L,B,L_N\) respectively. A typical result using the solution to Stechkin's problem is Theorem. If \(A\) is a homogeneous operator and \(Q\) is a centrally symmetric convex set, then for \(\delta >0\) \[ \omega(\delta)\leq E_\delta({\mathbb O}) \leq E_\delta({\mathbb B}) = E_\delta({\mathbb L}) \leq \ell(\delta). \] If there exist an element \(x\in Q\) and an operator \(T\in {\mathbb B}\) with the property \((1)\) then we have the equalities \[ \|Ax\|= \omega(\delta)= E_\delta({\mathbb O}) = E_\delta({\mathbb L}), \delta=\|x\|. \] The above inequalities are complemented and sharpened using the notion of Chebyshev radius and center of a bounded set in a normed space. The author also states a more general problem of recovery, namely the problem of recovering the values of the operator \(A\) on elements of \(Q\) from an exact or approximate value \(Ix\) of some information operator \(I\) on the element \(x\). This problem can be stated as follows: Recovery problem. Let \(X\) be some set, \(Y, Z\) metric spaces, \(A: X\to Y\), with domain \(D(A)\subset X\), \(Q\subset D(A)\), \(I: Q\to Z\) and \(\mathbb R\) a family of single valued maps from \(Z\) to \(Y\). For a number \(\delta\geq 0\) and an operator \(T\in {\mathbb R}\) the quantity \[ U_\delta(T)= \sup\{\|Ax-T\eta\|_Y \bigm|x\in Q,\eta\in Z, \|Ix-\eta\|_Z\leq \delta\} \] is the error of recovering the operator \(A\) on the elements of the class \(Q\) given exactly (if \(\delta=0\)) or approximately (\(\delta>0\)) by the information operator \(I\) using the method \(T\). The quantity of best recovery is then \( \inf \{U_\delta) \bigm|T\in{\mathbb R}\}\). \S 3. The approximation of functionals has special characteristics. For example, for considerably general classes of functionals \(A\), sets \(Q\) and approximating functionals \({\mathbb R}\), there necessarily exists a linear functional of best approximation. In particular \[ E_\delta({\mathbb O})= E_\delta(X^*)=\omega(\delta). \] Two of the three theorems cited in this section are Theorem 3.1 and Theorem 3.2. Theorem {3.1}. Let \(A\) be a linear functional with domain \(D(A) \subset X\), and let \(Q\) be a centrally symmetric convex subset of \(D(A)\). Then for any \(N>0\) \[ E(N)=\sup\{\omega(\delta) -N\delta \bigm|\delta>0\}. \] and an extremal functional exists in problem \((0.2)\). Theorem 3.2. Under the assumptions of Theorem 3.1 we have \[ E_\delta({\mathbb O})= E_\delta(X^*)=\ell(\delta)=\omega(\delta), \quad\delta>0. \] and if \(T_0\) is an extremal functional in problem \((2.2)\), then it will also be extremal in problem \((0.2)\), in which \(N\) is a subdifferential (derivative) of the function \(\omega(\delta)\) at the corresponding point. \S 4. One of the most studied problems, at the present time, is Stechkin problem on the best approximation of the differential operator \(D^k\) of order \(k\) on the class \[ Q=Q^n_{r,p}= \{x\in L_r \bigm|x^{(n)}\in L_p, \|x^{(n)}\|_{L_p} \leq 1\} \] of \(n\)-times differentiable functions \((0\leq k<n)\), \[ E(N)=E_{n,k}(N)=\inf\{U(T) \bigm|\|T\|_{L_r\to L_q} \leq N\}, \] \[ U(T)=\sup\{\|x^{(k)}-Tx\|_{L_q} \bigm|x\in Q^n_{r,p}\}. \] The spaces \(L_\gamma=L_\gamma(S), 1\leq\gamma\leq\infty\) are considered on the real line \(S=R\) or the half-line \(S=R_+\). \(L_\infty\) is the space of bounded continuous functions if \(q=\infty\) or \(r=\infty\) and is the space of measurable, essentially bounded functions if \(p=\infty\). The corresponding \(\omega(\delta)\) function is defined as \[ \omega(\delta)=\sup\{\|x^{(k)}\|_{L_q} \bigm|\|x\|_{L_r}\leq\delta, \|x^{(n)}\|_{L_p}\leq 1\}. \] If \(\mu=n-1/p+1/r>0\) (which is assumed in most theorems), then one verifies that \[ \omega(\delta)=K\delta^\alpha, K=\omega(1), \alpha=(n-k-1/p+1/q)/ (n-1/p+1/r). \tag{4.1} \] \(K\) is the smallest (best) constant in an inequality between the norms of the derivatives of functions (Kolmogorov inequality) \[ \|x^{(k)} \|_{L_q}\leq K\|x\|^\alpha_{L_r} \|x^{(n)}\|^\beta_{L_p}, \beta=1-\alpha.\tag{4.2} \] A result due to Stechkin is that if \(k+1/r-1/q>0\) then \[ E(N)\geq \beta\alpha^{\alpha/\beta}K^{1/\beta}N^{-\alpha/\beta}, N>0, \] for the best constant \(K\) above. Further theorems in this section give necessary and sufficient conditions, in certain cases, for \(K\) to be finite and for \(E_{n,k}(N)\) to be finite. Another result gives various conditions on \(p,q,r,k\) and \(n\) for the lower bound \(E(N)\) to be strict. The equality question in this lower bound has not been resolved as yet. A large number of papers are cited in this section dealing with particular cases in which the best constant \(K\) has been determined. Finally the author formulates a corresponding ill-posed problem of optimal recovery of values of the differential operator \(D^k\) of order \(k\) on the class \(Q=Q^n_{r,p}\) on \(n\)-times differentiable functions which amounts to the best numerical differentiation of functions defined with an error. Results in the directions discussed above are stated with regards to this problem. \S 5. The author brings several cases of exact results of approximation of differentiation operators on classes of functions of two variables. He also mentions a number of general results related to Stechkin problem on classes of functions of several variables belonging to Sobolev spaces and the corresponding related multiplicative inequalities between the norms of the derivatives in the spaces \(L_p(S), 0<p\leq\infty\) for \(S=R^m\) and \(S=R^m_+\). Here the Sobolev seminorm \[ \|f\|_{L^n_p(S)} = \max\{\|D^if\|_{L_P(S)} \bigm||i|=n\} \] is used where \[ D^kf= {\partial^{|k|}f \over \partial t_1^{k_1}\ldots\partial t_m^{k_m} } , |k|=k_1+\ldots k_m, f(t)=f(t_1,\ldots,t_m), k=(k_1,\ldots,k_m), k_i\in Z_+. \] No detailed theorems are formulated. \S 6. If in Stechkin's problem both the operator \(A\) and the class \(Q\) are invariant relative to a family of tranformations, then under additional conditions one can restrict oneself to invariant operators \(T\) in \(E(N)\) and \(E_\delta({\mathbb R})\) in(0.2) and (2.2). For example if \(T\) is a bounded linear operator in the space \(C_{2\pi}\) of continuous \(2\pi\)-periodic functions, then the operator \(\widetilde T\) defined by \[ \widetilde Tf={1\over 2\pi}\int_0^{2\pi}\tau_{-h}T\tau_h dh \] is linear, translation invariant, \(\|\widetilde T\|\leq\|T\|\) and so \(\widetilde T\) approximates a translation invariant class of functions no worse than \(T\). The averaging operation was used in approximation of a differentiable operator, particularly differential operators with constant coefficients which are translation invariant, and in the study of \(E(N)\) as \(n\to\infty\). Below we describe one general instance in which the invariance idea plays an essential role in simplifying the approximation problems described in earlier sections. Assume that there exist a semigroup \(S\) (with multiplicative operation) and families \(\Theta=\{\theta(s), s\in S\}, {\mathbb T}=\{\tau(s), s\in S\}\) of bounded linear operators in \(X\) and \(Y\) respectively which are themselves semigroups that is \[ \theta(st)=\theta(s)\theta(t), \tau(st)=\tau(s)\tau(t), t,s,\in S, \] such that for \(s\in S\) \[ \theta(s)Q\subset Q, \tau(s)A\theta(s)=A, \|\theta(s)\|_{X\to X}\leq 1, \|\tau(s)\|_{Y\to Y}\leq 1. \] Denote by \({\mathbb R}(S)\) sets of invariant operators belonging to \({\mathbb R}\) that is operators \(T\in {\mathbb R}\) with the property \(\tau(s)T\theta(s)=T\) on \(X\) for all \(s\in S\). Then under additional assumption in the basic Stechkin problems there is an extremal operator that possesses the invariance property and in particular \[ E(N)=\inf\{U(T) \bigm|T \in{\mathbb L}_N(S)\}, \quad E_\delta({\mathbb L})= E_\delta({\mathbb L}(S)). \] The complete conditions of several applications in this section are two long to state here and are omitted. \S 7. In this last section the author mentions a few results concerning the approximation of one class of differentiable functions by another smoother class of differentiable functions. Let \(p',r'.q', 1\leq p',r',q'\leq\infty; \ell,n\) integers, \(0\leq\ell\leq n\). Let \(B_{r'}^\ell(N)\) the set of functions \(\varphi\) with \(\|\varphi^{(n)}\|_{L^{r'}}\leq N; B^\ell_{q'}=B^\ell_{q'}(1)\). For \(\varphi\in B^\ell_{q'}\) put \[ F(\psi,N)= \inf\{\|\psi-\varphi\|_{L^{p'}} \bigm|\varphi\in B^n_{r'}(N)\} \] and let the quantity \[ F(N) =\sup\{ F(\psi,N) \bigm|\psi\in B^\ell_{q'} \} \] be representing the degree of best approximation in \(L_{p'}(-\infty,\infty)\) of the class \(B^\ell_{q'}\) by the class \(B^n_{r'}(N)\). The corresponding degree of linear approximation of a class by a class will be denoted by \(G(N)\). It is known that \(F(N)=N^{-\gamma}F(1), G(N)=N^{-\gamma}G(1)\) where \(\gamma=(\ell+1/p'-1/q')/(n-\ell+1/q'-1/r')\). \(F(N)\) is finite if and only if \(\ell/r'+(n-\ell)/p'\leq n/q'\). The quantities \(F(N), G(N)\) were found for concrete values of the parameters and sufficient conditions for the inequalities \[ F(N)\leq G(N) \leq E(N) \] were also given. A general result of a duality character is given by Theorem. Let \(1\leq p,q,r\leq\infty\), \[ \ell=n-k, 1/p+1/p'=1, 1/r+1/r'=1, 1/q+1/q'=1 \] and assume that \[ {n-k\over r}+{k\over p} \geq {n \over q} \] where \(q>r\) if \(k=0\). Then for any \(N>0\) we have the equality \[ F(N)=\beta\alpha^{\alpha/\beta}K^{1/\beta}N^{-\alpha/\beta} \] where \(K\) is the smallest constant in \((4.1)\) and \(\alpha\) and \(\beta\) are defined by \((4.1)\) and \((4.2)\). The inequality assumed in the last theorem is the finiteness condition for the constant \(K\) and the quantity \(F(N)\). Lastly we quote the Theorem. If the first assumption on the constants in the last theorem holds and in addition \(1<q,p\leq\infty, 1\leq r\leq\infty, \) then for any \(N>0\) \[ G(N)=E(N). \] An extensive bibliography of one hundred thirty three citations, eighteen of which are associated with the author and mostly by Russian mathematicians closely associated with Stechkin, is included in this article.

Recommendations

Cited in

(85)

This page was built for publication: Approximation of unbounded operators by bounded operators and related extremal problems

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4375223)