Finite sections of some classical inequalities

From MaRDI portal





The reviewer could not give a more illuminating idea of the contents of this slim little volume than that given by the author in his Introduction, as follows: ``We consider any classical inequality of the form \[ f(x_1,x_2, \ldots) < A g(x_1,x_2, \ldots), \tag{1} \] where the \(x_i\) are real variables and where \(A\) is a best possible constant; that is, if \(A\) were replaced by any smaller number there would exist a set of \(x\) for which (1) would be false. If we restrict (1) to just \(n\) variables, we find that \[ f(x_1,x_2,\ldots,x_n,0,0,\ldots) < A g(x_1,x_2,\ldots,x_n,0,0,\ldots) \] which is surely true, but in general will no longer be best possible for each \(n\). Indeed, we should expect to find a sequence of constants \(A_n\) \((n=1,2,\ldots)\) such that \[ f(x_1,x_2,\ldots,x_n,0,0,\ldots) \le A_n g(x_1,x_2,\ldots,x_n,0,0,\ldots) \tag{2} \] and \[ A - A_n \rightarrow 0\quad (n\rightarrow \infty). \tag{3} \] The object of this volume may now be concisely stated: to investigate refinements of the assertion (3) and to elucidate some of the theory which enters into such investigations. This theory includes, at one point or another, functional analysis, orthogonal polynomials, potential theory, Dirichlet series, number theory and other branches of analysis. It forms a coherent and deep theoretical structure, still in a stage of vigorous development, which makes use of and itself illuminates substantial bodies of related mathematics. Numerous areas remain unclear, as we shall try to point out along the way, and these insure the continuing development of the subject in the future. By way of explanation of the title of this volume, the inequality (2) is called a finite section of the inequality (1). Some parts of the book deal with finite sections \[ \int_0^a \int_0^a f(x,y)\,dx\,dy < A_n \int_0^a g(x)\,dx, \] or \[ \int_{-a}^a \int_{-a}^a f(x,y)\,dx \,dy < A_n \int_{-a}^a g(x)\,dx \] of the corresponding integral inequalities over unbounded intervals. In the first chapter the author presents some of the classical results on which the field is based. These results fall into three main categories and we shall discuss each of these, together with the corresponding related results in the rest of the volume. The first category of classical results deals with Toeplitz forms associated with functions \(f\) which are real, periodic of period \(2\pi\), and in \(L(-\pi,\pi)\). The \(n\)-th Toeplitz form associated with \(f\) is the quadratic form \[ (x,A_nx)_n = \sum_{\mu, \nu =0}^n a_{\mu - \nu} x_\mu \bar x_\nu, \] where \[ a_n = (2\pi)^{-1} \int_{-\pi}^\pi e^{-int} f(t)\,dt, \] and the Toeplitz matrix associated with \(f\) is the infinite matrix \(A = \left(a_{\mu - \nu}\right)\), whose sections are the \((n +1) \times (n +1)\) Hermitian matrices \(A\). The central question in the classical theory of Toeplitz forms is the relationship between the function \(f\) and the distribution of the spectrum of \(A\). If \(m\) and \(M\) are the essential infimum and supremum of \(f\) on \((-\pi,\pi)\), and if \(\lambda_1^{(n)}\le \lambda_2^{(n)}\le \cdots\le \lambda_{n+1}^{(n)}\) are the \(n+1\) eigenvalues of the matrix \(A_n\), then \[ \lambda_1^{(n)}(x, x)_n \le \lambda_{n+1}^{(n)}(x, x)_n \] are the (well-known) finite sections of the inequalities \(m(x,x) \le (x,Ax) \le M(x,x)\). For a complete treatment of the basic theory, the reader is referred to the book ``Toeplitz forms and their applications by \textit{U. Grenander} and \textit{G. Szegő}[Berkeley: University of California Press (1958; Zbl 0080.09501)]. In the spirit of this volume, the main question is the behaviour of \(\lambda_\nu^{(n)}\) as \(n\rightarrow\infty\). In Chapter II, rather than dealing with this question for Toeplitz matrices, the author considers the corresponding question for Toeplitz kernels. Here the natural starting point is the integral equation of Toeplitz type: \[ \lambda\varphi(x) = \int_{-a}^a K(x - y)\varphi(y)\,dy \] whose kernel \(K\) is real, even, and in \(L( -\infty, \infty)\). If the eigenvalues are denoted by \(\lambda_{1,a}\ge \lambda_{2,a}\ge \cdots\), then the main question now concerns the behaviour of \(\lambda_{\nu,a}\) as \(a\to \infty\), and the relationship between the \(\lambda_{\nu,a}\) and the Fourier transform \(F(t) = \int_{-\infty}^\infty e^{itx} K(x)\,dx\). Using classical function-theoretic methods of \textit{H. Widom} [Trans. Am. Math. Soc. 88, 491--522 (1958; Zbl 0101.09202)] it is shown that under appropriate hypotheses on \(F\) [including \(M = \max F(t) = F(0)\) and \(\sigma^2 = - F(0) \ne 0] \) \[ \lambda_{\nu,a} = M - (1/8) \sigma^2 \nu^2 \pi^2 a^{-2} + O(a^{-2}) \] as \(a\to \infty\). (Widom (loc. cit.) had proved somewhat more than this, and had obtained the corresponding results for Toeplitz matrices.) The author next gives a generalization of these results -- also due to \textit{H. Widom} [Trans. Am. Math. Soc. 106, 391--414 (1963; Zbl 0205.14603)] -- to certain \(N\)-dimensional convolution operators, using functional analysis. As an application of Widom's first theorem, the author shows that the best possible constant \(\lambda_n\), for the finite section \[ \sum_{\mu, \nu=1}^n (\mu+\nu)^{-1} x_\mu x_\nu \le \lambda_n \sum_{\mu =1}^n x_\mu^2 \] of Hilbert's inequality (with \(n=\infty\) and \(\lambda_\infty = \pi)\) has the form \[ \lambda_n \equiv \pi - \frac12 \pi^5 (\log n)^{-2} + O(\log \log n/(\log n)^3) \] as \(n\to\infty\). The same method is applied to the more general inequality of Schur which states that if \(K(u,v)\) is a nonnegative, symmetric kernel on \(I\times I\), with \(I = (0, \infty)\), which is homogeneous of degree \(-1\) and such that \(u^{-1/2} K(u,1)\) and \(v^{-1/2} K(1,v)\) are strictly decreasing, then \[ \sum_{\mu, \nu=1}^\infty K(\mu,\nu) x_\mu x_\nu < M\, \sum_{\mu=1}^\infty x_\mu^2,\quad\text{unless }x_\mu \equiv 0, \] and \[ \int_0^\infty \int_0^\infty K(u,v)f(u)f(v)\,du\,dv < M\, \int_0^\infty f^2(u)\,du,\quad\text{unless } f = 0, \] where \(M = \int_0^\infty u^{-1/2} K(u,1)\,du\) is best possible. Proofs of these results are given in Chapter I, while the same simple transformation of the integral equation \( \Lambda_0 f(u) = \int_1^n K(u,v) f(v)\,dv\) \((n= e^{2n})\) into one of Toeplitz type as that used for Hilbert's inequality gives \[ \lambda_n = M - \gamma\pi^2 (\log n)^{-2} + O(\log \log n/(\log n)^3) \] as the best possible constant for the finite section \[ \sum_{\mu,\nu=1}^n K(\mu,\nu) x_\mu x_\nu \le \lambda_n \sum_{\mu=1}^n x_\mu^2,\quad (\gamma= \int_1^\infty t^{-1/2} (\log t)^2 K(t,1)\,dt) \] of Schur's inequality, provided \(K(1,t)=O(t^{-1/2-\delta})\) as \(t\to\infty\). (The result for the discrete case is deduced from the result for integrals.) Chapter II concludes with a brief discussion of the dual relationship between Toeplitz kernels and kernels \(K\) which are symmetric and homogeneous of degree \(-1\). \bigskip The second category of classical results is concerned with Hankel forms, and Hankel matrices \(H=(\mu_{i+j})\) in which the entries depend only on the sum \((i+j)\) of the subscripts. The Hilbert matrix, appearing in Hilbert's inequality above, has \(\mu_{i+j}=(i+j)^{-1}\) and is thus a Hankel matrix. In case the matrix \(H\) arises from the moments, \(\mu_r = \int_a^b x^r w(x)\,dx\) \((r =0,1,\ldots)\), of a nonnegative weight function \(w\) on \([a,b]\), then \(H\) is positive definite. Hence if \(\{\lambda_N\}_0^\infty\) is the sequence of smallest eigenvalues of the \(N\)-th finite sections \(H_N= (\mu_{i+j})_{i,j=0}^N\) of \(H\), then \(\lambda_N>0\) for all \(N\ge 0\). The question of the rate of approach of the \(\lambda_N\) to zero is reduced to a question of the asympotic rate of growth of a sequence \(\{P_n(x)\}\) of polynomials which are orthogonal with respect to the weight function \(w\). The details are rather technical in nature and are given in Chapter III (Hankel forms), with appropriate references to the monograph ``Orthogonal Polynomials by \textit{G. Szegő} [Colloq. Publ. Vol. 23, New York: AMS (1959; Zbl 0089.27501)]. The remainder of Chapter III may be summarized in the author's own words, as follows: ``Following the discussion of the lowest eigenvalue of \(H_N\) we turn to the largest eigenvalue, giving a theorem of Nehari which yields the precise conditions under which the spectra of the sections of Hankel matrices are uniformly bounded. We then discuss a theorem of Widom which settles a problem raised in the previous chapter on the spectral density of Hankel matrices in general, and the Hilbert matrix in particular. The chapter concludes with a survey of related results and generalizations. As an example of the results of this chapter, it is found that the lowest eigenvalue of the Hilbert matrix \(((m +n +1)^{-1})_{m,n=0}^N\) is \[ \lambda_N \sim \pi^{3/2} (8 + 6\sqrt 2)^{5/2}N^{1/2} (1 + \sqrt 2)^{-(4N+9)}\quad\text{as }N\to\infty. \] The third category of classical results is more restricted in nature, and deals with (only a few of) those inequalities which do not assert the boundedness of some linear operator on \(\ell^2\). The inequalities in question are Carleman's inequality: if all \(x_n\ge 0\), then \[ \sum_1^\infty (x_1x_2\cdots x_n)^{1/n} <e \sum_1^\infty x_n\quad\text{unless }x_n\equiv 0; \] and Hardy's inequality: if \(p>1\) and all \(x_n\ge 0\), then \[ \sum_1^\infty \left(\frac{x_1+x_2+\ldots +x_n}{n}\right)^p < (1 - p^{-1})^{-p} \sum_1^\infty x_n^p \quad\text{unless }x_n\equiv 0. \] (Both of these inequalities have integral analogues not referred to by Wilf; these appear as Theorems 327 and 335 in \textit{G. H. Hardy}, \textit{J. E. Littlewood}, and \textit{G. Pólya} [Inequalities. 2nd ed., Cambridge: At the University Press (1952; Zbl 0047.05302)].) Carleman's inequality is proved in Chapter I (Basic Results), while in Chapter IV (Nonlinear Theory) the author gives the proof of \textit{N. G. de Bruijn} [Nederl. Akad. Wet., Proc., Ser. A 66, 505--514 (1963; Zbl 0115.27702)] that the best constant for the \(n\)-th section of Carleman's inequality is \[ \lambda_n =e - 2\pi^2 e (\log n)^{-2} + O\left((\log n)^{-3}\right) \text{ as }n\to\infty. \] Wilf refers to similar results for Hardy's inequality (and others) in a thesis of \textit{Stan T. M. Ackermans} [An asymptotic method in the theory of series. Ph. D. Diss. Eindhoven (1964; \url{http://alexandria.tue.nl/extra3/proefschrift/PRF1A/7706595.pdf})]. There is a generalization of Schur's inequality due to Hardy, Littlewood, and Pólya (loc. cit., Theorem 318, and the integral analogue, Theorem 319) which states that if \(p>1\) and \(q = p/(p -1)\), while \(K(x,y)\) is a nonnegative kernel which is homogeneous of degree \(-1\) and such that \(x^{-1/p}K(x,1)\) and \(y^{-1/q}K(1,y)\) are strictly decreasing, then if all \(x_m\ge 0\) and \(y_n \ge 0\), \[ \sum_{n=1}^\infty \left(\sum_{m=1}^\infty K(m,n) x_m\right)^p < k^p\, \sum_{n=1}^\infty x_n^p \quad\text{unless }x_n\equiv 0, \] and \[ \sum_{m,n=1}^\infty K(m,n) x_m y_n < k\, \left(\sum x_m^p\right)^{1/p}\left(\sum y_n^q\right)^{1/q}\quad\text{unless }x_m\equiv 0\text{ or }y_n\equiv 0, \] where \(k = \int_0^\infty x^{-1/p} K(x,1)\,dx\). Wilf notes that, at present, no general theory is known which gives the rate of convergence of the best possible constant \(\lambda_n^p\) in the corresponding finite sections \[ \sum_{r=1}^n \left( \sum_{s=1}^n K(s,r) x_s\right)^p \le \lambda_n^p \sum_{r=1}^n x_r^p \tag{4} \] to its limit \(k^p\), and regards this as one of the major unsolved problems of the subject. In this final chapter, Wilf has taken an important step toward such a general theory by proving the following minimax principle: if \(p>1\), and \(K_n\), is an \(n\times n\) matrix with nonnegative entries such that \(K'_nK_n\) has strictly positive entries, while \[ \lambda_n = \max \{\Vert K_nx\Vert_p: x= (x_1, \ldots, x_n) \ge 0,\ \Vert x\Vert_p^p = \sum_1^n x_\nu^p = 1\} \] is the constant appearing in (4), then \[ \lambda_n^p = \min_{x>0} \max_{1\le r\le n} x_r^{1- p} \sum_{\nu=1}^n K(r,\nu) \left(\sum_{\mu=1}^n K(\mu,\nu)x_\mu\right)^{p-1}. \] \bigskip \{Reviewer's remarks: This valuable monograph is written in a clear, comprehensible manner, and should stimulate further research on such inequalities. Very few typographical or other errors were detected by the reviewer. The only misleading ones occur on p. 7 where the condition ``\(p>0\) for (1.19) should be ``\(p>1\), and where the hypotheses for Theorem 1.3 should be changed to those stated above, preceding (4). Although the following references do not strictly deal with finite sections of infinite inequalities, nevertheless it would have been useful if the author had referred to results of this kind, if only by way of contrast: \textit{A. M. Pfeffer} [J. Res. Natl. Bur. Stand., Sect. B 70, 221--231 (1966; Zbl 0145.06101)], \textit{R. Redheffer} [Proc. Lond. Math. Soc. (3) 17, 683--699 (1967; Zbl 0156.05804)], \textit{D. C. Benson} [J. Math. Anal. Appl. 17, 292--308 (1967; Zbl 0146.07404)], the reviewer [Can. J. Math. 21, 222--234 (1969; Zbl 0165.37001)].\}




Cited in
(25)








This page was built for publication: Finite sections of some classical inequalities

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q5597673)