Convergence of the method of least squares

Convergence of probability measures (60B10) Probabilistic methods, stochastic differential equations (65C99) Numerical smoothing, curve fitting (65D10)

Review text

Let \(f(t)\) be defined and continuous on the interval \([-1,1]\). Suppose we are given measurements \(y_ j = f(t_ j) + g_ j\) \((j = 0,1,\dots,N- 1)\) of \(f\), where \(\{t_ 1,t_ 2,\dots,t_{N - 1}\}\) is a system of distinct points on the segment \([-1,1]\); and \(g_ j\) \((j = 0,1,\dots,N- 1)\) is the observational error, which is an independent random variable satisfying the conditions \(E(g_ j) = 0\), \(E(g_ kg_ j) = \sigma \delta_{jk}\). Let \(S_{n,N}(t)\) denote an algebraic polynomial minimizing the sum \(\sum^{N-1}_{j = 0} (y_ j - P_ n(t_ j))^ 2\) over all the polynomials \(P_ n(t)\) of degree \(n \leq N-1\). If \(J_{n,N}(f,t) = E((f(t) - S_{n,N}(t))^ 2)\) then the following result is obtained. Theorem: \(\text{If }f\in \text{Lip}_ s M\) \((s \geq 1/2)\), the condition \(n^ 2/N \leq a_ n = o(1)\) \((N \to \infty)\) is necessary and sufficient for \(\max_{-1 \leq t \leq 1} J_{n,N} (f,t)\) to tend to zero as \(n \to \infty\). If the function \(f\) satisfies a Dini-Lipschitz condition on \([-1,1]\), then the condition \(n^ 2/N \leq a\) \((a>0)\) is sufficient for \(\max\{J_{n,N}(f,t): -1 + \varepsilon \leq t \leq 1 - \varepsilon\}\) to approach zero as \(n \to \infty\).

Recommendations

Cites work

Cited in

(8)

This page was built for publication: Convergence of the method of least squares

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q1326060)