John R. Kender, Rick Kjeldsen
IEEE Transactions on Pattern Analysis and Machine Intelligence
We explicitly obtain, for K(x, y) totally positive, a best choice of functions u1, ..., un and v1, ..., vn for the problem minui, vi (∝01 (∝01 |K(x, y) - ∑i = 1, n ui(x) vi(y)| dyp dx) 1 p, where ui ε{lunate} Lp[0, 1], vi ε{lunate} L1[0, 1], i = 1, ..., n, and p ε{lunate} [1, ∞]. We show that an optimal choice is determined by certain sections K(x, ξ1), ..., K(x, ξn), and K(τ1, y), ..., K(τn, y) of the kernel K. We also determine the n-widths, both in the sense of Kolmogorov and of Gel'fand, and identify optimal subspaces, for the set Kr,v = {f(x) = ∑ i=1 raiki(x) + ∫ 0 1K(x,y)h(y)dy, (a1, ..., ar)ε{lunate}Rr, {norm of matrix}h{norm of matrix}p≤1}, as a subset of Lq[0, 1], with either p = ∞ and q ε{lunate} [1, ∞], or p ε{lunate} [1, ∞] and q = 1, where {k1(x), ..., kr(x), K(x, y)} satisfy certain restrictions. A particular example is the ball Br,v = {f} in the Sobolev space. © 1978.
John R. Kender, Rick Kjeldsen
IEEE Transactions on Pattern Analysis and Machine Intelligence
Heng Cao, Haifeng Xi, et al.
WSC 2003
Sonia Cafieri, Jon Lee, et al.
Journal of Global Optimization
Zhihua Xiong, Yixin Xu, et al.
International Journal of Modelling, Identification and Control