Abstract Generated abstract
This paper generalizes V. A. Markov’s inequalities for derivatives of algebraic polynomials on the interval [-1, 1] under a weighted majorization condition involving |alpha x + i sqrt(1 - x^2)|. It proves sharp upper bounds for all derivatives of a polynomial of degree at most n, expressed through derivatives at 1 of an extremal combination of Chebyshev polynomials, with equality only for constant unimodular multiples of this extremal polynomial at the endpoints. The proof uses previously established derivative estimates, interlacing properties of associated polynomial zeros, and positivity of even power series expansions related to Chebyshev functions, including a result of Schaeffer and Duffin. The result is also connected to problems on constructive characterization of functions by polynomial approximation.
Full Text
MATHEMATICS
V. S. VIDENSKII
A GENERALIZATION OF V. A. MARKOV’S INEQUALITIES
(Presented by Academician S. N. Bernstein on 14 I 1958)
Theorem. If on the interval \([-1,1]\) a polynomial \(P_n(x)\) of degree not exceeding \(n\) satisfies the inequality
\[ |P_n(x)| \leq |\alpha x+i\sqrt{1-x^2}| \quad (\alpha>0), \tag{1} \]
then
\[ |P_n^{(k)}(x)| \leq M_n^{(k)}(1)=\frac{\alpha+1}{2}T_n^{(k)}(1)+\frac{\alpha-1}{2}T_{n-2}^{(k)}(1) \quad (k=1,\ldots,n), \tag{2} \]
where \(T_n(x)=\cos n\arccos x\). Equality in (2) is attained only for polynomials \(P_n(x)=\gamma M_n(x)\), \(|\gamma|=1\),
\[ M_n(x)=\frac{\alpha+1}{2}T_n(x)+\frac{\alpha-1}{2}T_{n-2}(x) \tag{3} \]
at the points \(x=\pm1\).
For \(\alpha=1\) the right-hand side of (1) is equal to unity, and inequalities (2) become V. A. Markov’s inequalities. For \(k=1\) and \(2\) and any \(\alpha>0\) the theorem was proved in my note \((^1)\).
Put
\[ M_n(x)=\Re\{(\alpha x+i\sqrt{1-x^2})[T_{n-1}(x)+iS_{n-1}(x)]\}, \]
\[ L_n(x)=\sqrt{1-x^2}N_{n-1}(x)=\Im\{(\alpha x+i\sqrt{1-x^2})[T_{n-1}(x)+iS_{n-1}(x)]\}, \tag{4} \]
where \(S_n(x)=\sin n\arccos x\). The functions \(M_n(x)\) and \(N_{n-1}(x)\) are polynomials of degrees \(n\) and \(n-1\), respectively; all their zeros lie in the interval \((-1,1)\) and mutually interlace (see, for example, \((^2)\)). It is not difficult to show that from (4) there follow equality (3) and the equality
\[ L_n(x)=\frac{\alpha+1}{2}S_n(x)+\frac{\alpha-1}{2}S_{n-2}(x). \tag{5} \]
If we put
\[ H_k(x)=|M_n^{(k)}(x)+iL_n^{(k)}(x)| \quad (k=1,\ldots,n); \tag{6} \]
\[ \Phi_k(x)= \begin{cases} H_k(x), & \text{for } \xi_1^{(k)} \leq x \leq \xi_{n-k+1}^{(k)},\\ |M_n^{(k)}(x)|, & \text{for } -\infty < x \leq \xi_1^{(k)},\ \xi_{n-k+1}^{(k)} \leq x < +\infty, \end{cases} \tag{7} \]
where \(\xi_1^{(k)}\) and \(\xi_{n-k+1}^{(k)}\) are the extreme zeros of the function \(L_n^{(k)}(x)\) lying on the interval \([-1,1]\), then from my previous results \((^{3-5})\) it follows that, for polynomials satisfying (1), the following estimates of their derivatives are valid:
\[ |P_n^{(k)}(x)| \leq \Phi_k(x) \quad (k=1,\ldots,n;\ -\infty<x<\infty). \tag{8} \]
It was originally shown by S. N. Bernstein \((^6)\) that on the interval \([-1,1]\) for the first derivative the inequality \(|P'_n(x)| \leqslant H_1(x)\) holds.
Our theorem will follow directly from inequality (8), if we show that the even continuous function \(\Phi_k(x)\) increases monotonically for \(x>0\). Moreover, since all zeros of \(M_n^{(k)}(x)\) lie in the interval \((\xi_1^{(k)},\xi_{n-k+1}^{(k)})\) \((^3)\), it is necessary to establish the increase of \(\Phi_k(x)\) only on the interval \([0,\xi_{n-k+1}^{(k)}]\).
Let us prove that \(H_k^2(x)\) is expanded in the interval \((-1,1)\) in a Taylor series in even powers of \(x\) with positive coefficients. In doing so we shall rely on the fact established in the work of A. Schaeffer and R. Duffin \((^7)\) that
\[ W_{n,k}(x)=[T_n^{(k)}(x)]^2+[S_n^{(k)}(x)]^2=\sum_{p=0}^{\infty} a_{p,k}x^{2p},\qquad a_{p,k}>0 \tag{9} \]
\[ (k=1,2,\ldots,n;\ p=0,1,2,\ldots). \]
Obviously, the function \(H_k^2(x)\) can be written in the form
\[ \begin{aligned} H_k^2(x)=& \left\{\frac{\alpha}{2}[T_n^{(k)}(x)+T_{n-2}^{(k)}(x)] +\frac{1}{2}[T_n^{(k)}(x)-T_{n-2}^{(k)}(x)]\right\}^2 \\ &+\left\{\frac{\alpha}{2}[S_n^{(k)}(x)+S_{n-2}^{(k)}(x)] +\frac{1}{2}[S_n^{(k)}(x)-S_{n-2}^{(k)}(x)]\right\}^2 . \end{aligned} \tag{10} \]
On the one hand, the identities
\[ T_n(x)+T_{n-2}(x)=2xT_{n-1}(x),\qquad T_n(x)-T_{n-2}(x)=-\frac{2}{n-1}(1-x^2)T'_{n-1}(x), \]
\[ S_n(x)+S_{n-2}(x)=2xS_{n-1}(x),\qquad S_n(x)-S_{n-2}(x)=-\frac{2}{n-1}(1-x^2)S'_{n-1}(x). \tag{11} \]
hold.
On the other hand, the functions \(T_n(x)\) and \(S_n(x)\) satisfy the differential equation
\[ (1-x^2)y''-xy'+n^2y=0, \tag{12} \]
therefore \(T_n^{(k)}(x)\) and \(S_n^{(k)}(x)\) satisfy the equation
\[ (1-x^2)y^{(k+2)}-(2k+1)xy^{(k+1)}+(n^2-k^2)y^{(k)}=0. \tag{13} \]
From the identities (11) we obtain
\[ [T_n^{(k)}(x)+T_{n-2}^{(k)}(x)]^2+[S_n^{(k)}(x)+S_{n-2}^{(k)}(x)]^2= \]
\[ =4\left[x^2W_{n-1,k}(x)+kx\frac{d}{dx}W_{n-1,k-1}(x)+k^2W_{n-1,k-1}(x)\right]. \tag{14} \]
From (13) we obtain
\[ [(1-x^2)T'_{n-1}(x)]^{(k)} =-[xT_{n-1}^{(k)}(x)+(n^2-2n+k)T_{n-1}^{(k-1)}(x)], \]
\[ [(1-x^2)S'_{n-1}(x)]^{(k)} =-[xS_{n-1}^{(k)}(x)+(n^2-2n+k)S_{n-1}^{(k-1)}(x)]. \tag{15} \]
Applying (11) and (15), we may write
\[ [T_n^{(k)}(x)-T_{n-2}^{(k)}(x)]^2+[S_n^{(k)}(x)-S_{n-2}^{(k)}(x)]^2= \]
\[ =\frac{4}{(n-1)^2}\left[ x^2W_{n-1,k}(x)+(n^2-2n+k)x\frac{d}{dx}W_{n-1,k-1}(x)+ \right. \]
\[ \left. +(n^2-2n+k)^2W_{n-1,k-1}(x) \right]; \tag{16} \]
\[
\{[T_n^{(k)}(x)]^2-[T_{n-2}^{(k)}(x)]^2\}+\{[S_n^{(k)}(x)]^2-[S_{n-2}^{(k)}(x)]^2\} =
\]
\[
= \frac{2}{n-1}\left[2x^2 W_{n-1,k}(x)+(n^2-2n+2k)x\frac{d}{dx}W_{n-1,k-1}(x)+\right.
\]
\[
\left.{}+2k(n^2-2n+k)W_{n-1,k-1}(x)\right].
\tag{17}
\]
From (14), (16), and (17), taking (9) into account, we conclude that for \(k=1,\ldots,n-2\) the function \(H_k^2(x)\) expands on \((-1,1)\) into an even power series with positive coefficients, and hence for \(1\le k\le n-2\) the theorem is proved. For \(k=n\) the theorem follows directly from (8), since at \(x=1\) we have \(|P_n^{(n)}(1)|\le M_n^{(n)}(1)\). For \(k=n-1\) the result again follows easily from (8), for the maximum of the linear function \(P_n^{(n-1)}(x)\) on the interval \([-1,1]\) is attained at one of the endpoints of the interval; but at the points \(x=\pm1\) we have \(|P_n^{(n-1)}(\pm1)|\le M_n^{(n-1)}(1)\). In the proof of the theorem it was implicitly assumed that \(n\ge2\), but it is clear that for \(n=1\) and \(2\) inequalities (2) are simple consequences of inequalities (8).
I note that, in the question of a constructive characterization of functions given on a finite interval by means of their approximations by polynomials, investigated recently by V. K. Dzyadyk \({}^{8}\) and A. F. Timan \({}^{9}\), an important role is played by estimates of the successive derivatives of polynomials satisfying the inequalities
\[ |P_n(x)|\le \left|\frac{x}{n}+i\sqrt{1-x^2}\right|^\rho,\qquad \rho>0,\qquad -1\le x\le 1. \]
In the works mentioned, neither an extremal polynomial for this problem, for any values of \(\rho\), nor exact upper bounds for the derivatives \(P_n^{(k)}(x)\) on the interval \([-1,1]\) were given.
Mathematical Institute named after V. A. Steklov
Academy of Sciences of the USSR
Received
14 I 1958
CITED LITERATURE
\({}^{1}\) V. S. Videnskii, Zap. matem. otd. fiz.-matem. fak. i Kharkovsk. matem. obshch., 26, ser. 4 (1958).
\({}^{2}\) S. N. Bernstein, Collected Works, 1, article No. 42, 1952, pp. 452—467.
\({}^{3}\) V. S. Videnskii, DAN, 67, No. 5 (1949).
\({}^{4}\) V. S. Videnskii, DAN, 73, No. 2 (1950).
\({}^{5}\) V. S. Videnskii, Izv. AN SSSR, ser. matem., 15, 401 (1951).
\({}^{6}\) S. N. Bernstein, Collected Works, 1, article No. 46, 1952, pp. 497—499.
\({}^{7}\) C. A. Schaeffer, R. J. Duffin, Bull. Am. Math. Soc., 44, No. 4 (1938).
\({}^{8}\) V. K. Dzyadyk, Izv. AN SSSR, ser. matem., 20, 623 (1956).
\({}^{9}\) A. F. Timan, DAN, 116, No. 5 (1957).