No label defined (Q5285645)

The problem considered is that of justification of the Bayesian approach to nonlinear regression. The simplest of such models looks as follows: \[ y_ i = g(t_ i) + \varepsilon_ i, \;t_ i \in [0,1],\;i=1, \dots,n, \] with the random vector \((\varepsilon_ 1, \dots, \varepsilon_ n) \sim N (0, \sigma^ 2I_{n \times n}),\) the constant \(\sigma^ 2\) is unknown and \(g(\cdot)\) is an unknown function with \(m-1\) continuous derivatives and \(\int^ 1_ 0 (g^{(m)} (t))^ 2dt < \infty\). Let \(W_ 2^{(m)}\) be the set of all such functions. It was shown by \textit{I. J. Schoenberg} [Proc. Natl. Acad. Sci. USA 52, 947-950 (1964; Zbl 0147.321)], that the solution of the extreme problem \[ \arg \min_{g \in W_ 2^{(m)}} \left( n^{-1} \sum^ n_{j=1} \bigl( g(t_ i) -y_ i \bigr)^ 2 + \lambda \int^ 1_ 0 \bigl( g^{(m)} (t) \bigr)^ 2dt \right) =\;? \] is a polynomial spline of degree \(2m- 1\). Within the framework of the Bayesian approach it was suggested by \textit{G. Wahba} [J. R. Stat. Soc., Ser. B 40, 364-372 (1978; Zbl 0407.62048)] to take as a prior distribution for \(g(\cdot)\) that of the process \[ x(t) = \sum^ m_{j=1} \theta_ jt^{j-1}/(j-1)! + b^{1/2} z(t), \] where the random vector \((\theta_ 1, \dots, \theta_ m) \sim N(0,\xi I_{m \times m})\), \(\xi<0\), and \(z(t)\) is an \(m\)-fold integrated Wiener process, i.e. \[ z(t) = \int^ t_ 0 (t-u)^{m-1}/(m-1)! dW(u). \] However, the prior measure is supported by the functional set which does not contain \(W^{(m)}_ 2\). In order to avoid this obstacle the author suggests to define a prior distribution as that of an abstract Wiener process [see \textit{H. Kuo}, ``Gaussian measures in Banach spaces'' (1975; Zbl 0306.28010)].