On a Class of Interpolation Polynomials with Unfixed Nodes
Unknown
Submitted 1965-01-01 | SovietRxiv: ru-196501.14193 | Translated from Russian

Abstract Generated abstract

This paper generalizes a theorem of Mycielski and Paszkowski on interpolation polynomials with unfixed nodes, replacing algebraic polynomials by generalized polynomials generated by systems of functions satisfying Chebyshev-type zero-counting conditions. For prescribed positive interpolation values with alternating signs, it studies the existence and uniqueness of interior nodes at which the derivative of the corresponding interpolation polynomial vanishes. The proof develops determinant representations for fundamental interpolation functions and uses sign changes, zero-counting arguments, and induction on the number of unfixed nodes. The main result establishes, for systems in the class \(T_n'\), the unique choice of any prescribed block of interior nodes so that the associated generalized interpolation polynomial has zero derivative at all nodes in that block.

Full Text

MATHEMATICS

V. S. VIDENSKII

ON A CLASS OF INTERPOLATION POLYNOMIALS WITH UNFIXED NODES

(Presented by Academician S. N. Bernstein on 2 XII 1964)

In the work of Ya. Mycielski and S. Paszkowski \((^1)\) the following interesting theorem was proved. Let \(v_0, v_1, \ldots, v_n\) be given positive numbers and let \([a,b]\) be a real interval. There exists a unique sequence of points
\[ a=x_0<x_1<\cdots<x_{n-1}<x_n=b \]
and such an algebraic polynomial \(P_n(x)\) of degree \(n\) that
\[ P_n(x_k)=(-1)^{\,n-k}v_k \quad (k=0,1,\ldots,n); \]
\[ P_n'(x_i)=0 \quad (i=1,\ldots,n-1). \]
In particular, if all \(v_k=1\), then \(P_n(x)\) is the Chebyshev polynomial. Here we generalize this theorem and give a proof based on an entirely different idea than in \((^1)\). Consider a sequence of functions \(\{\varphi_k(x)\}_{k=0}^n\), continuously differentiable on \([a,b]\). We shall write \(\{\varphi_k(x)\}\in T_n\) if every polynomial
\[ P(x)=\sum_{k=0}^n a_k\varphi_k(x) \quad \left(\sum_{k=0}^n |a_k|>0\right) \]
on \([a,b]\) has \(\le n\) zeros. If, moreover, its derivative \(P'(x)\) also has \(\le n\) zeros on \([a,b]\), then we shall write \(\{\varphi_k(x)\}\in T_n'\).

Fix positive numbers \(v_0,v_1,\ldots,v_n\) and, having chosen a system of interpolation nodes
\[ a=\xi_0<\xi_1<\cdots<\xi_{n-1}<\xi_n=b, \]
construct, with respect to \(\{\varphi_k(x)\}\), the polynomial
\[ P(x)=P(x;\xi_0,\xi_1,\ldots,\xi_n)=\sum_{k=0}^n (-1)^{\,n-k}v_k l_k(x), \tag{1} \]
where
\[ l_k(x)=D_k(x)/D[\varphi_0(\xi_0),\varphi_1(\xi_1),\ldots,\varphi_n(\xi_n)], \]
\[ D[\varphi_0(\xi_0),\varphi_1(\xi_1),\ldots,\varphi_n(\xi_n)] = \begin{vmatrix} \varphi_0(\xi_0) & \varphi_1(\xi_0) & \cdots & \varphi_n(\xi_0)\\ \varphi_0(\xi_1) & \varphi_1(\xi_1) & \cdots & \varphi_n(\xi_1)\\ \cdot & \cdot & \cdot & \cdot\\ \varphi_0(\xi_n) & \varphi_1(\xi_n) & \cdots & \varphi_n(\xi_n) \end{vmatrix}, \]
\[ D_k(x)=D[\varphi_0(\xi_0),\ldots,\varphi_{k-1}(\xi_{k-1}),\varphi_k(x),\varphi_{k+1}(\xi_{k+1}),\ldots,\varphi_n(\xi_n)]. \]
The class of polynomials of the form (1), for fixed \(v_k>0\) and arbitrary \(\xi_1<\cdots<\xi_{n-1}\) from \((a,b)\), will be denoted by \(V_n\).

Theorem 1. Let \(\{\varphi_k(x)\}\in T_n\), and suppose that the points
\[ a=\xi_0<\cdots<\xi_{i-1}<\xi_{i+1}<\cdots<\xi_n=b \quad (1\le i\le n-1) \]
and the numbers \(w_0,\ldots,w_{i-1},w_i,w_{i+1},\ldots,w_n\) are given such that
\[ w_{i-1}>0,\quad w_i\le 0,\quad w_{i+1}>0. \]
In \((\xi_{i-1},\xi_{i+1})\) there exists such a point \(\xi_i\) and such a polynomial \(L(x)\) with respect to the system \(\{\varphi_k(x)\}\) that
\[ L(\xi_k)=w_k \quad (k=0,1,\ldots,n);\qquad L'(\xi_i)=0. \]

Adjoin to the points \(\xi_k\) \((k\ne i)\), as an interpolation node, an arbitrary point \(z\) from \((\xi_{i-1},\xi_{i+1})\), and construct the polynomial
\[ L(x)=L(x;\xi_0,\ldots,\xi_{i-1},z,\xi_{i+1},\ldots,\xi_n) =\sum_{k=0}^n w_k l_k(x), \tag{2} \]

where \(l_k(x)\) are the fundamental polynomials for the system of nodes
\(\xi_0<\cdots<\xi_{i-1}<z<\xi_{i+1}<\cdots<\xi_n\). Differentiating formula (2) with respect to \(x\) at the point \(x=z\), we obtain

\[ L'(z)D_i(z)=\sum_{k=0}^{n} w_kD_k'(z)=\Phi(z),\quad \text{where } \quad D_k'(z)=D_k'(x)\big|_{x=z}. \]

We shall show that \(\Phi(\xi_{i-1})\Phi(\xi_{i+1})<0\). Consequently, there is a point \(z=\xi_i,\ \xi_{i-1}<\xi_i<\xi_{i+1}\), such that \(\Phi(\xi_i)=0\), and hence also \(L'(\xi_i)=0\), since \(D_i(\xi_i)=D[\varphi_0(\xi_0),\ldots,\varphi_n(\xi_n)]\ne0\) for \(\{\varphi_k(x)\}\in T_n\). If \(i\ne k\), then
\(D_k'(z)=D[\varphi_0(\xi_0),\ldots,\varphi_i(z),\ldots,\varphi_k'(z),\ldots,\varphi_n(\xi_n)]\). For \(i\ne k\) the determinant \(D_k'(\xi_j)=0\), if \(j\ne i,k\), since it contains two identical rows. Thus,

\[ \Phi(\xi_{i-1})=w_{i-1}D_{i-1}'(\xi_{i-1})+w_iD_i'(\xi_{i-1}),\quad \Phi(\xi_{i+1})=w_iD_i'(\xi_{i+1})+w_{i+1}D_{i+1}'(\xi_{i+1}) \]

The polynomial \(D_i(x)\) vanishes at the \(n\) points \(\xi_j\) \((j\ne i)\); hence the \(\xi_j\) are its simple zeros, and therefore, in particular, \(D_i'(\xi_{i-1})\ne0\). Since the determinant \(D_{i-1}'(\xi_{i-1})\) is obtained from the determinant \(D_i'(\xi_{i-1})\) by interchanging two of its rows, we have
\(D_{i-1}'(\xi_{i-1})=-D_i'(\xi_{i-1})\). We obtain

\[ \Phi(\xi_{i-1})=(-w_{i-1}+w_i)D_i'(\xi_{i-1}),\quad \Phi(\xi_{i+1})=(w_i-w_{i+1})D_i'(\xi_{i+1}), \]

and since \(\xi_{i-1},\xi_{i+1}\) are neighboring simple zeros of the polynomial \(D_i(x)\), it follows that
\(D_i'(\xi_{i-1})D_i'(\xi_{i+1})<0\), and consequently
\(\Phi(\xi_{i-1})\Phi(\xi_{i+1})<0\).

Theorem 2. Let \(\{\varphi_k(x)\}\in T_n'\), and let points be given

\[ a=\xi_0<\xi_1<\cdots<\xi_{i-1}<\xi_{i+r}<\cdots<\xi_{n-1}<\xi_n=b \]
\[ (1\le i\le n-1;\quad 1\le r\le n-1;\quad i+r\le n). \]

In \((\xi_{i-1},\xi_{i+r})\) there exists a unique system of \(r\) points
\(\xi_i<\xi_{i+1}<\cdots<\xi_{i+r-1}\) such that
\(P(x)=P(x;\xi_0,\ldots,\xi_n)\in V_n\) satisfies the conditions

\[ P'(\xi_i)=P'(\xi_{i+1})=\cdots=P'(\xi_{i+r-1})=0. \tag{3} \]

Theorem 2 generalizes the result of [1] in two directions: algebraic polynomials are replaced by polynomials with respect to \(\{\varphi_k(x)\}\in T_n'\), and instead of \(r=n-1\) the condition \(1\le r\le n-1\) is introduced.

We shall prove Theorem 2 by induction applied to the number \(r\). For \(r=1\) the result follows from Theorem 1. Indeed, suppose that in \((\xi_{i-1},\xi_{i+1})\) there exist two points \(\xi_{i1}\) and \(\xi_{i2}\) such that the corresponding polynomials \(P_1(x)\in V_n\) and \(P_2(x)\in V_n\) satisfy the conditions
\(P_1'(\xi_{i1})=0,\ P_2'(\xi_{i2})=0\). Since \(\xi_{i1}\) and \(\xi_{i2}\) are the unique points of maximum of the polynomials \((-1)^{\,n-i}P_1(x)\), \((-1)^{\,n-i}P_2(x)\) in \((\xi_{i-1},\xi_{i+1})\), it follows that
\(v_i=|P_1(\xi_{i1})|>(-1)^{\,n-i}P_2(\xi_{i1})\),
\((-1)^{\,n-i}P_1(\xi_{i2})<|P_2(\xi_{i2})|=v_i\). Hence it follows that the polynomial \(P_1(x)-P_2(x)\) has a zero in \((\xi_{i1},\xi_{i2})\). On the other hand,
\(P_1(\xi_k)-P_2(\xi_k)=0\) \((k\ne i)\); consequently, the total number of zeros of \(P_1(x)-P_2(x)\) is greater than \(n\) on \([a,b]\).

We shall now assume that Theorem 2 has already been established for some fixed number \(r\) \((1\le r\le n-2)\) for any \(i\), and shall show that in this case it is valid also for the number \(r+1\). Thus, let points be given
\(a=\xi_0<\xi_1<\cdots<\xi_{i-1}<\xi_{i+r+1}<\cdots<\xi_n=b\). Adjoin to them an arbitrary point
\(\xi_{i+r}'\), \(\xi_{i-1}<\xi_{i+r}'<\xi_{i+r+1}\). By the induction hypothesis, in
\((\xi_{i-1},\xi_{i+r}')\) there are uniquely found \(r\) points
\(\xi_i'<\xi_{i+1}'<\cdots<\xi_{i+r-1}'\) such that the polynomial

\[ P_1(x)=P_1(x;\xi_0,\ldots,\xi_{i-1},\xi_i',\ldots,\xi_{i+r}',\xi_{i+r+1},\ldots,\xi_n)\in V_n \]

satisfies (3) at these points. If \(P_1'(\xi_{i+r}')=0\), then \(P_1(x)\) is the desired polynomial. If, however, \(P_1'(\xi_{i+r}')\ne0\), then in
\((\xi_{i+r-1}',\xi_{i+r+1})\) there is a point
\(\eta_{i+r}'\), \(\eta_{i+r}'\ne\xi_{i+r}'\), such that
\(P_1(\eta_{i+r}')=(-1)^{\,n-i-r}v_{i+r}\). Next choose

\[ \xi_{i+r}''=\frac12(\xi_{i+r}'+\eta_{i+r}') \]

and repeat the preceding reasoning. In

\((\xi_{i-1}, \xi''_{i+r})\) there are uniquely determined \(r\) points \(\xi''_i < \cdots < \xi''_{i+r-1}\) such that the polynomial

\[ P_2(x)=P_2(x;\xi_0,\ldots,\xi_{i-1},\xi''_i,\ldots,\xi''_{i+r},\xi_{i+r+1},\ldots,\xi_n)\in V_n \]

satisfies conditions (3) at these points. Suppose that \(P_2'(\xi''_{i+r})\ne 0\); denote by \(x'_1<\cdots<x'_n\) the zeros of \(P_1(x)\), and let us show that \(P_1(x)-P_2(x)\ne 0\) in \((x'_{i+r},x'_{i+r+1})\). Indeed, if at some point \(z\), \(x'_{i+r}<z<x'_{i+r+1}\),

\[ P_1(z)=P_2(z)=(-1)^{n-i-r}v^*_{i+r}, \qquad v^*_{i+r}>0, \]

then this would mean that the two distinct polynomials

\[ P_1(x)=P_1(x;\xi_0,\ldots,\xi'_{i+r-1},z,\xi_{i+r+1},\ldots,\xi_n)\in V_n^*, \]

\[ P_2(x)=P_2(x;\xi_0,\ldots,\xi''_{i+r-1},z,\xi_{i+r+1},\ldots,\xi_n)\in V_n^* \]

satisfy conditions (3), where \(V_n^*\) is the class of polynomials obtained from \(V_n\) by replacing \(v_{i+r}\) by \(v^*_{i+r}\). But this contradicts the induction hypothesis on the uniqueness of such a polynomial. Since \(P_1(x)-P_2(x)\ne 0\) in \((\eta'_{i+r},\xi'_{i+r})\subset (x'_{i+r},x'_{i+r+1})\), and since \(P_1(\xi'_{i+r})=P_1(\eta'_{i+r})=P_2(\xi'_{i+r})\), it follows that \(P_2(x)<|P_1(x)|\) for \(x\in[\eta'_{i+r},\xi'_{i+r}]\), and in \((\eta'_{i+r},\xi'_{i+r})\) there is a point \(\eta''_{i+r}\ne \xi'_{i+r}\) such that

\[ P_2(\eta''_{i+r})=(-1)^{n-i-r}v_{i+r}. \]

It is clear that \(|\eta''_{i+r}-\xi''_{i+r}|<\frac12|\eta'_{i+r}-\xi'_{i+r}|\). We shall now show that

\[ \xi_i<\xi''_i,\ldots,\xi'_{i+r-1}<\xi''_{i+r-1}. \tag{4} \]

We shall regard the polynomial \(P_2(x)\) as a continuous image of the polynomial \(P_1(x)\) corresponding to the continuous displacement of the point \(\xi'_{i+r}\) to \(\xi''_{i+r}\). Move the point \(\xi'_{i+r}\) to the point \(\widetilde{\xi}_{i+r}\), and assume that \(|\xi'_{i+r}-\widetilde{\xi}_{i+r}|\) is as small as we need. In \((\xi_{i-1},\widetilde{\xi}_{i+r})\) there are uniquely determined \(r\) points \(\widetilde{\xi}_i<\cdots<\widetilde{\xi}_{i+r-1}\), which are, respectively, small displacements of the points \(\xi'_i<\cdots<\xi'_{i+r-1}\), such that the polynomial

\[ Q'(x)=Q'(x;\xi_0,\ldots,\xi_{i-1},\widetilde{\xi}_i,\ldots,\widetilde{\xi}_{i+r},\ldots,\xi_n)\in V_n' \]

satisfies conditions (3) at the points \(\widetilde{\xi}_i,\ldots,\widetilde{\xi}_{i+r-1}\). If \(|\widetilde{\xi}_k-\xi'_k|\) \((k=i,\ldots,i+r-1)\) are sufficiently small, then

\[ \xi_i<\widetilde{\xi}_i,\ldots,\xi'_{i+r-1}<\widetilde{\xi}_{i+r-1}. \tag{5} \]

Indeed, if for some \(k\) we had \(\widetilde{\xi}_k<\xi'_k\), then the polynomial \(P_1(x)-Q'(x)\) would have two zeros in a neighborhood of the point \(\xi'_k\). In a neighborhood of each of the \(\xi'_j\) it has at least one zero, so that the total number of its zeros in \((\xi_{i-1},\xi_{i+r+1})\) would be \(\ge r+1\), while at the remaining \(n-r\) nodes \(\xi_s\), \(P_1(\xi_s)-Q'(\xi_s)=0\); hence on \([a,b]\) the function \(P_1(x)-Q'(x)\) would have \(\ge n+1\) zeros. This proves also inequalities (4), since inequalities (5) cannot be violated as \(\xi'_{i+r}\to \xi''_{i+r}\).

Continuing the construction process by the same scheme, we obtain \(\{P_m(x)\}\subset V_n\). We see that \(\{\xi_k^{(m)}\}\) \((k=i,\ldots,i+r-1)\), for fixed \(k\), is increasing and bounded; consequently, as \(m\to\infty\) it has a limit \(\xi_k\); \(\{\xi_{i+r}^{(m)}\}\) has a limit as \(m\to\infty\), since

\[ [\eta_{i+r}^{(m)},\xi_{i+r}^{(m)}]\subset[\eta_{i+r}^{(m-1)},\xi_{i+r}^{(m-1)}] \]

and

\[ |\eta_{i+r}^{(m)}-\xi_{i+r}^{(m)}|<2^{-m+1}|\eta'_{i+r}-\xi'_{i+r}|. \]

Let us note that \(\xi_k\ne \xi_{k+1}\). Indeed, if \(\xi_k=\xi_{k+1}\), then \(\{P_m(x)\}\) would converge to a discontinuous function; meanwhile, \(\{P_m(x)\}\) is uniformly bounded on \([\xi'_{i+r-1},\xi'_{i+r}]\), hence it is compact and can have as its limit only a polynomial from \(T_n'\). Put

\[ P(x)=\lim_{m\to\infty}P_m(x). \]

It is easy to see that \(P(x)=P(x;\xi_0,\ldots,\xi_n)\in V_n\) and that

\[ P'(\xi_i)=\cdots=P'(\xi_{i+r})=0. \tag{6} \]

We shall now show that the points \(\xi_i<\cdots<\xi_{i+r}\) and the corresponding polynomial \(P(x)\in V_n\) are uniquely determined by conditions (6). Suppose that there exists another system of points \(\zeta_i<\cdots<\zeta_{i+r}\) such that the polynomial \(Q(x)\in V_n\) corresponding to it satisfies conditions (6) at these points. If \(\xi_{i+r}=\zeta_{i+r}\), then for the two systems of \(r\) points \(\xi_i<\cdots<\xi_{i+r-1}\) and \(\zeta_i<\cdots<\zeta_{i+r-1}\), and for the polynomials \(P(x)\) and \(Q(x)\), conditions (3) would be fulfilled, which contradicts the inductive assumption of uniqueness. Let, for definiteness, \(\xi_{i+r}<\zeta_{i+r}\). We begin our construction process by choosing in \((\xi_{i-1},\xi_{i+r+1})\) an arbitrary point \(\xi'_{i+r}\). Denote by \(X\) the upper bound of those \(\xi'_{i+r}\) starting from which we arrive at the points \(\xi_i<\cdots<\xi_{i+r}\) and at the polynomial \(P(x)\). It is clear that \(X\le \zeta_{i+r}\), since the sequence of contracting intervals \([\eta^{(m)}_{i+r},\xi^{(m)}_{i+r}]\) is constructed uniquely after the point \(\xi'_{i+r}\) has been chosen. Suppose that \(X=\zeta_{i+r}\). Choose \(\xi'_{i+r}<X\) and so close to \(X\) that the polynomial \(Q(x)\) does not change sign in \((\xi'_{i+r},X)\) and so that \(\xi_{i+r}<\xi'_{i+r}\). Then \(\xi'_{i+r}\) will be the right endpoint of the interval \([\eta'_{i+r},\xi'_{i+r}]\), and the corresponding polynomial \(P_1(x)\) will be monotone on \([\xi'_{i+r},\xi_{i+r+1}]\). Hence it follows that \(|P_1(X)|<|Q(X)|,\ |P_1(\xi'_{i+r})|>|Q(\xi'_{i+r})|\). Consequently, in \((\xi'_{i+r},X)\) there is a point \(\mu_{i+r}\) such that
\(P_1(\mu_{i+r})=Q(\mu_{i+r})=(-1)^{\,n-i-r}v^*_{i+r},\ v^*_{i+r}>0\), and this means that two distinct polynomials

\[ \begin{aligned} P_1(x)&=P_1(x;\xi_0,\ldots,\xi_{i+r-1},\mu_{i+r},\ldots,\xi_n)\in V_n^*,\\ Q(x)&=Q(x;\xi_0,\ldots,\xi_{i+r-1},\mu_{i+r},\ldots,\xi_n)\in V_n^* \end{aligned} \]

satisfy conditions (3) at the points \(\xi_i<\cdots<\xi_{i+r-1}\) and \(\zeta_i<\cdots<\zeta_{i+r-1}\), respectively; \(V_n^*\) is the class obtained from \(V_n\) by replacing \(v_{i+r}\) by \(v^*_{i+r}\). We have arrived at a contradiction with the inductive assumption of the uniqueness of such a polynomial. The case \(X<\zeta_{i+r}\) is considered analogously. It is only necessary, in addition to the point \(\xi'_{i+r}<X\), to choose another point \(\zeta'_{i+r}>X\) and, with the aid of our process, to construct the polynomial \(Q_1(x)\), and then repeat arguments similar to those given above.

Leningrad Electrotechnical Institute of Communications
named after M. A. Bonch-Bruevich

Received
1 XII 1964

REFERENCES

  1. J. Mycielski, S. Paszkowski, Bull. Acad. Polon. Sci., Ser. math., astronom., phys., 8, No. 7, 433 (1960).

Submission history

On a Class of Interpolation Polynomials with Unfixed Nodes