On Some Fundamental Inequalities in the Theory of Orthogonal Polynomials
Unknown
Submitted 1961-01-01 | SovietRxiv: ru-196101.70287 | Translated from Russian

Abstract Generated abstract

This paper studies local inequalities for orthonormal polynomials on the unit circle, relating positivity of the derivative of the orthogonality distribution, growth bounds for the Christoffel kernel, and pointwise boundedness of the polynomials. It proves that, under a local regularity condition, a linear bound on the kernel implies positivity of the density, while pointwise boundedness implies the kernel bound, and it analyzes why the converse boundedness question can fail through the recurrence parameters of the circle polynomials. Further results establish almost everywhere equivalences on arcs, existence of bounded subsequences, and boundedness of the full sequence under a condition on the recurrence parameters, with corresponding consequences for orthogonal polynomials on an interval.

Full Text

MATHEMATICS

Ya. L. Geronimus

ON SOME BASIC INEQUALITIES IN THE THEORY OF ORTHOGONAL POLYNOMIALS

(Presented by Academician S. N. Bernstein, 27 V 1961)

1. Let the polynomials \(\{\varphi_n(z)\}_0^\infty\) be orthonormal on the circle \(z=e^{i\theta}\) \((0\leq \theta \leq 2\pi)\) with respect to the distribution \(d\sigma(\theta)\); consider three inequalities

\[ \sigma'(\theta_0)\geq m>0; \tag{I} \]

\[ K_n(z_0)=\sum_{i=0}^{n}|\varphi_i(z_0)|^2\leq C(n+1),\qquad z_0=e^{i\theta_0}; \tag{II} \]

\[ \{|\varphi_n(z_0)|\}_0^\alpha \leq M. \tag{III} \]

Theorem 1. If \(\sigma'(\theta_0)\) exists and if

\[ \int_0^h |dt\{\sigma(\theta_0+t)-\sigma(\theta_0-t)-2t\sigma'(\theta_0)\}|=o(h), \tag{1} \]

then (I) follows from (II).

The proof follows from the inequality \((n+1)\leq K_n(z_0)\sigma_n(\theta_0)\), where \(\sigma_n(\theta)\) is the Fejér sum for the Fourier–Stieltjes series \(\mathfrak{S}(d\sigma)\); from condition (1) there follows the convergence \(\lim_{n\to\infty}\sigma_n(\theta_0)=\sigma'(\theta_0)\).

Theorem 1 is also valid in the more general case when (II) holds only for a subsequence \(\{K_{n_\nu}(z_0)\}\), and also in the case when there exists only the generalized derivative

\[ \sigma^{(1)}(\theta_0)=\lim_{h\to 0}\frac{\sigma(\theta_0+h)-\sigma(\theta_0-h)}{2h}. \tag{2} \]

We see that (II) implies (I); it is also obvious that (III) implies (II), and consequently also (I),—but does (III) follow from (I) or from (II)? This question was posed in its time by V. A. Steklov (1); in our works (2, 3) we obtained some sufficient conditions for the boundedness of the whole orthonormal system on the entire interval of orthogonality or on a part of it, but in doing so we had to impose on the function \(\sigma(\theta)\) certain restrictions not only of a local but also of a global character, and the question remained open.

2. We shall use the relations

\[ \sqrt{1-|a_n|^2}\,\varphi_{n+1}(z)=z\varphi_n(z)-\bar a_n\varphi_n^*(z), \]

\[ \sqrt{1-|a_n|^2}\,\varphi_{n+1}^*(z)=\varphi_n^*(z)-a_n z\varphi_n(z), \]

\[ \varphi_n^*(z)=z^n\overline{\varphi_n\!\left(\frac{1}{z}\right)} \qquad (n=0,1,\ldots), \tag{3} \]

where the parameters \(\{a_n\}_0^\infty\) are subject to the conditions \(\{|a_n|\}_0^\infty<1\); conversely, if the parameters are chosen completely arbitrarily and independently of one another

1002

one another, subjecting them only to these conditions, then the polynomials \(\{\varphi_n(z)\}_0^\infty\), constructed by the recurrence formula (3), are orthonormal on \(|z|=1\) with respect to a certain distribution \(d\sigma(\theta)\), which is completely determined by these parameters. From (3) there follows the formula

\[ \sqrt{1-|a_n|^2}=\frac{2\Re\eta_n}{1+|\eta_n|^2}\leqslant \frac{2|\eta_n|}{1+|\eta_n|^2},\qquad \eta_n=\frac{z_0\varphi_n(z_0)}{\varphi_{n+1}(z_0)},\qquad z_0=e^{i\theta_0}. \tag{4} \]

Theorem 2. Under condition (II) one has \(\varliminf_{n\to\infty}|\varphi_n(z_0)|<\infty\); if at the same time \(\varlimsup_{n\to\infty}|\varphi_n(z_0)|=\infty\), then \(\lim_{n\to\infty}|a_n|=1\).

If we had \(\lim_{n\to\infty}|\varphi_n(z_0)|=\infty\), then from this would follow the existence of the limits \(\lim_{n\to\infty}|\varphi_n(z_0)|=\infty\) and \(\lim_{n\to\infty} K_n(z_0)/(n+1)=\infty\), which contradicts condition (II).

Suppose that \(\varlimsup_{n\to\infty}|\varphi_n(z_0)|=\infty\); we divide the set of all natural numbers into two sequences \(\{n_i'\}\) and \(\{n_k''\}\) so as to have

\[ |\varphi_{n_i'}(z_0)|\leqslant A\quad (i=1,2,\ldots);\qquad \lim_{k\to\infty}|\varphi_{n_k''}(z_0)|=\infty; \tag{5} \]

both sequences are, obviously, infinite.

We shall consider pairs of numbers \(n\) and \(n+1\) belonging to two different sequences—such pairs will always be found for arbitrarily large values of \(n\), since, if beginning with \(n=n_0\) no such pairs were found, then one of the sequences would have no more than \(n_0\) terms. Choose \(n\) so large that \(|\eta_n|<\varepsilon\) or \(|\eta_n|>1/\varepsilon\), where \(\varepsilon>0\) is arbitrarily small; in both cases we obtain from (2) \(\sqrt{1-|a_n|^2}<2\varepsilon\), which is equivalent to the condition \(\lim_{n\to\infty}|a_n|=1\).

Remark 1. Formula (4) shows that, knowing \(\varphi_n(z_0)\), we may choose \(\varphi_{n+1}(z_0)\ne0\) completely arbitrarily—this determines \(|a_n|\), and by formula (3) also \(a_n\); under condition (II) we shall indicate several examples of such a choice:

1) Put \(\varphi_m(z_0)=C\sqrt{m}\) for \(m=2^k\) \((k=0,1,2,\ldots)\) and \(\varphi_m(z_0)=1\) for all other values of \(m\).

2) Let \(\varphi_m(z_0)=C\sqrt[3]{m}\) for \(m=k^3\) \((k=1,2,3,\ldots)\) and \(\varphi_m(z_0)=1\) for all other values of \(m\); it can be shown that in both cases we have

\[ \lim_{n\to\infty}\left[\prod_{k=0}^{n-1}(1-|a_k|^2)\right]^{1/2}=1, \]

whence it follows that the set of points of increase of the function \(\sigma(\theta)\) is everywhere dense on the interval \([0,2\pi]\).

  1. Theorem 3. 1) From the validity of (II) almost everywhere on \([\alpha,\beta]\) there follows the validity of (I) almost everywhere inside \([\alpha,\beta]\).

2) From the validity of (I) almost everywhere on \([\alpha,\beta]\) there follows the validity of (II) inside \([\alpha,\beta]\).

The proof of assertion 1) follows from the convergence \(\lim_{n\to\infty}\sigma_n(\theta)=\sigma'(\theta)\) almost everywhere on \([\alpha,\beta]\); assertion 2) is known.

Theorem 4. From the validity of (I) almost everywhere on \([\alpha,\beta]\) there follows the existence of an infinite subsequence of polynomials \(\{\varphi_{n_\nu}(e^{i\theta})\}\), bounded at every point inside \([\alpha,\beta]\), uniformly bounded on some interior subinterval; under the additional condition \(\varlimsup_{n\to\infty}|a_n|<1\) the entire sequence \(\{\varphi_n(e^{i\theta})\}_0^\infty\) has this property.

Indeed, from the proof of Theorem 2 it follows that, under the condition \(\overline{\lim}_{\nu\to\infty}|a_\nu|<1\), we have \(\overline{\lim}_{\nu\to\infty}|\varphi_\nu(e^{i\theta})|<\infty\) at every point at which (11) is satisfied; the uniform boundedness follows from Szegő’s theorem.

Remark 2. The condition \(\overline{\lim}_{n\to\infty}|a_n|<1\) is satisfied, in particular, if \(\lg \sigma'(\theta)\in L_1\), since this latter condition is equivalent to the condition
\[ \sum_{n=0}^{\infty}|a_n|^2<\infty; \]
but it is also satisfied in a number of other cases—for example, in the case considered by N. I. Akhiezer \((^4)\), since in this case there exists the limit \(\lim_{n\to\infty} a_n=a,\ |a|<1\).

  1. From the theorems proved there follow analogous theorems for polynomials \(\{p_n(x)\}_0^\infty\) orthonormal on the interval \([-1,+1]\) with respect to the distribution \(d\psi(x)\); here the condition \(\overline{\lim}_{n\to\infty}|a_n|<1\) is equivalent to the condition \(\underline{\lim}_{n\to\infty}\lambda_n>0\), where \(\lambda_n\) is the coefficient in the recurrence formula
    \[ \sqrt{\lambda_{n+1}}\,p_n(x)=(x-a_n)p_{n-1}(x)-\sqrt{\lambda_n}\,p_{n-2}(x)\qquad (n=1,2,\ldots); \tag{6} \]
    the condition \(\underline{\lim}_{n\to\infty}\lambda_n>0\) is satisfied, in particular, if \(\lg \psi'(x)/\sqrt{1-x^2}\in L_1\), and is also satisfied in a number of other cases, for example, in the case of Pollaczek polynomials, for which \(\lg \psi'(x)/\sqrt{1-x^2}\notin L_1\), but \(\lim_{n\to\infty}\lambda_n=1/4\).

Kharkov Aviation
Institute

Received
22 V 1961

REFERENCES CITED

\(^1\) V. A. Steklov, Izv. Rossiisk. Akad. nauk, 15, 281 (1921).
\(^2\) Ya. L. Geronimus, Izv. AN SSSR, ser. matem., 16, No. 5, 469 (1952).
\(^3\) Ya. L. Geronimus, Polynomials Orthogonal on the Circle and on an Interval, Moscow, 1958.
\(^4\) N. I. Akhiezer, DAN, 130, No. 2 (1960).

Submission history

On Some Fundamental Inequalities in the Theory of Orthogonal Polynomials