V. S. Ryaben’kii
Unknown
Submitted 1960-01-01 | SovietRxiv: ru-196001.33990 | Translated from Russian

Abstract Generated abstract

The paper studies interpolation and tabulation of periodic functions of several variables whose Fourier coefficients satisfy a product type algebraic decay condition. Using Korobov type lattice points on a parallelepipedal grid, it constructs approximate Fourier coefficients from only N function values and forms a trigonometric interpolation polynomial with frequencies restricted by a product cutoff. Lemmas on lattice quadrature, shifted Fourier coefficients, and divisor type sums yield an L2 error estimate with explicit dependence on N, the smoothness parameter alpha, the dimension s, and logarithmic factors. The paper also compares this construction with interpolation on a standard cubic grid, indicating that the parallelepipedal grid can avoid the deterioration with dimension exhibited by cubic grid tables for this function class.

Full Text

V. S. Ryaben’kii

ON TABLES AND INTERPOLATION OF FUNCTIONS FROM A CERTAIN CLASS

(Presented by Academician A. N. Kolmogorov, 20 XII 1959)

The class of functions \(E_s^\alpha(C)\) consists of all functions \(f(x_1,\ldots,x_s)\),

\[ f(x_1,\ldots,x_s)= \sum_{m_1,\ldots,m_s=-\infty}^{\infty} C(m_1,\ldots,m_s)e^{2\pi i(m_1x_1+\cdots+m_sx_s)}, \]

whose Fourier coefficients \(C(m_1,\ldots,m_s)\) satisfy the inequality

\[ |C(m_1,\ldots,m_s)| \leq \frac{C}{(\bar m_1\ldots \bar m_s)^\alpha}, \]

where \(\bar m_\nu=\max(|m_\nu|,1)\), \(\alpha>1\), and \(C\) does not depend on \(m_1,\ldots,m_s\).

Lemma 1. Let \(f\in E_s^\alpha(C)\). For every prime \(N>s\) one can specify integers \(a_1,\ldots,a_s\) such that

\[ \left| \int_0^1\cdots\int_0^1 f(x_1,\ldots,x_s)\,dx_1\ldots dx_s - \frac{1}{N}\sum_{k=1}^{N} f\left(\frac{ka_1}{N},\ldots,\frac{ka_s}{N}\right) \right| \leq A_0 C N^{-\alpha}\ln^{\alpha s}N, \]

where \(A_0=A_0(\alpha,s)\)*.

The proof is contained in the proof of Theorem 1 of the paper \((^1)\).

Lemma 2. If \(f\in E_s^\alpha(C)\), \(\bar n_1\ldots \bar n_s<N_1\), then

\[ \varphi(x_1,\ldots,x_s)\equiv f(x_1,\ldots,x_s)e^{-2\pi i(n_1x_1+\cdots+n_sx_s)} \in E_s^\alpha(A_1 C N_1^\alpha). \]

Proof. We note that

\[ \frac{\bar m}{\overline{(m-n)}\,\bar n}<A, \]

where \(A\) is a constant independent of \(m\) and \(n\). Let \(C(m_1,\ldots,m_s)\) and \(C'(m_1,\ldots,m_s)\) be the Fourier coefficients of the functions \(f\) and \(\varphi\), respectively. Then

\[ |C'(m_1,\ldots,m_s)| = |C(m_1-n_1,\ldots,m_s-n_s)| \leq \frac{C}{(\overline{m_1-n_1}\ldots \overline{m_s-n_s})^\alpha} = \]

\[ = \frac{C(\bar n_1\ldots \bar n_s)^\alpha}{(\bar m_1\ldots \bar m_s)^\alpha} \left[ \frac{\bar m_1}{\overline{(m_1-n_1)}\,\bar n_1} \cdots \frac{\bar m_s}{\overline{(m_s-n_s)}\,\bar n_s} \right]^\alpha \leq \frac{CA^{\alpha s}N_1^\alpha}{(\bar m_1\ldots \bar m_s)^\alpha}, \]

which proves the lemma.

* Here and below \(A_\nu=A_\nu(\alpha,s)\) denotes constants depending only on \(\alpha\) and \(s\).

Lemma 3. If \(1<N_1<N,\ \alpha>1\), then the estimates

\[ \sum_{\bar m_1\ldots \bar m_s\ge N_1} (\bar m_1\ldots \bar m_s)^{-\alpha} \ll A_2 N_1^{-(\alpha-1)}\ln^{s-1}N, \]

\[ \sum_{\bar m_1\ldots \bar m_s<N_1} 1 \ll A_3 N_1\ln^{s-1}N \]

hold.

Proof. The lemma is proved elementarily by induction on \(s\). Denote

\[ N_1=[\sqrt N\,\ln^{s/2}N], \]

\[ \widetilde C(m_1,\ldots,m_s) =\frac1N\sum_{k=1}^{N} f\left(\frac{ka_1}{N},\ldots,\frac{ka_s}{N}\right) e^{-2\pi i\,\frac{a_1m_1+\cdots+a_sm_s}{N}\,k} \tag{1} \]

\[ P(x_1,\ldots,x_s) = \sum_{\bar m_1\ldots \bar m_s<N_1} \widetilde C(m_1,\ldots,m_s) e^{2\pi i(m_1x_1+\cdots+m_sx_s)}. \]

The trigonometric polynomial (1) may be regarded as an interpolation polynomial for the function \(f(x_1,\ldots,x_s)\), constructed from the table of values of this function at the points
\[ \left(\frac{ka_1}{N},\ldots,\frac{ka_s}{N}\right),\quad k=1,2,\ldots,N. \]
The grid formed by these points in the space \(x_1,\ldots,x_s\) will be called parallelepipedal.

Theorem. If \(f\in E_s^\alpha(C)\), then the estimate

\[ \int_0^1\cdots\int_0^1 \left|f(x_1,\ldots,x_s)-P(x_1,\ldots,x_s)\right|^2 \,dx_1\cdots dx_s = O\!\left(N^{-\alpha+1/2}\ln^{(\alpha+1/2)s-1}N\right) \]

holds.

Proof. Applying abbreviated notation, we obtain

\[ \int |f(\mathbf x)-P(\mathbf x)|^2\,d\mathbf x = \sum_{\bar m_2\ldots \bar m_s<N_1} |C(\mathbf m)-\widetilde C(\mathbf m)|^2 + \sum_{\bar m_1\ldots \bar m_s\ge N_1} |C(\mathbf m)|^2, \tag{2} \]

where

\[ \widetilde C(\mathbf m)=\widetilde C(m_1,\ldots,m_s), \]

\[ C(\mathbf m)= \int_0^1\cdots\int_0^1 f(x_1,\ldots,x_s) e^{-2\pi i(x_1m_1+\cdots+x_sm_s)} \,dx_1\cdots dx_s. \]

By virtue of the definition of \(\widetilde C(\mathbf m)\), using Lemmas 1 and 2, we obtain

\[ |C(\mathbf m)-\widetilde C(\mathbf m)| \ll A_4 C N_1^\alpha N^{-\alpha}\ln^{\alpha s}N. \tag{3} \]

Observing that
\[ |C(\mathbf m)|^2\ll C^2(\bar m_1\ldots \bar m_s)^{-2\alpha}, \]
from (2) and (3), according to Lemma 3, we obtain

\[ \int |f(\mathbf x)-P(\mathbf x)|^2\,d\mathbf x \ll A_5C^2\left( N_1^{2\alpha+1}N^{-2\alpha}\ln^{2\alpha s+s-1}N + N_1^{-(2\alpha-1)}\ln^{s-1}N \right) = \]

\[ = O\!\left(N^{-(\alpha-1)/2}\ln^{(\alpha+1/2)s-1}N\right), \]

where the constant in the \(O\) depends on \(\alpha\), \(s\), and \(C\).

Let us compare the tables of values of functions \(f\in E_s^\alpha(C)\) at the points of a parallelepipedal grid with the tables of values of functions \(f\in E_s^\alpha(C)\) at the nodes of a cubic grid whose sides are parallel to the coordinate axes and have length \(h=1/n\). The number of all grid points lying in the unit cube,

i.e., \(N \cong n^s\). Any interpolation formula constructed from tables of values of functions \(f \in E_s^\alpha(C)\) at the nodes of a cubic grid has a remainder term of order no greater than \(O(N^{-2\alpha/s})\), which decreases with increasing \(s\). To prove this assertion it suffices to note that the function \(f_1(x_1,\ldots,x_s) \equiv 0\) and the function
\[ f_2(x_1,\ldots,x_s)=2C\frac{\sin n\pi x_1}{n^\alpha}\in E_s^\alpha(C) \]
have identical tables (identically zero), although
\[ \int |f_1(\mathbf{x})-f_2(\mathbf{x})|^2\,d\mathbf{x} =\frac{2C^2}{n^{2\alpha}} =O\left(N^{-\frac{2\alpha}{s}}\right). \]

After the completion of this work it became known to me that, independently of me, S. A. Smolyak was engaged in closely related questions (see \((^2)\)).

Received
10 XII 1959

CITED LITERATURE

\(^1\) N. M. Korobov, DAN, 124, No. 6, 1207 (1959).
\(^2\) S. A. Smolyak, DAN, 131, No. 5 (1960).

Submission history

V. S. Ryaben’kii