Abstract
In this paper, we consider the optimal design problem for extrapolation and estimation of the slope at a given point, say z, in a polynomial regression with no intercept. We provide explicit solutions of these problems in many cases and characterize those values of z, where this is not possible.
Similar content being viewed by others
References
Celant, G., Broniatowski, M. (2016). Interpolation and extrapolation designs for the polynomial regression, chapter 4, pp. 69–111. New York: Wiley.
Chang, F.-C., Heiligers, B. (1996). E-optimal designs for polynomial regression without intercept. Journal of Statistical Planning and Inference, 55(3), 371–387.
Chang, F.-C., Lin, G.-C. (1997). \(D\)-optimal designs for weighted polynomial regression. Journal of Statistical Planning and Inference, 62, 317–331.
Dette, H. (1990). A generalization of \(D\)- and \(D_1\)-optimal designs in polynomial regression. Annals of Statistics, 18, 1784–1805.
Dette, H., Franke, T. (2001). Robust designs for polynomial regression by maximizing a minimum of \(D-\) and \(D_1\)-efficiencies. Annals of Statistics, 29(4), 1024–1049.
Dette, H., Huang, M.-N. L. (2000). Convex optimal designs for compound polynomial extrapolation. Annals of the Institute of Statistical Mathematics, 52(3), 557–573.
Dette, H., Melas, V. B., Pepelyshev, A. (2004). Optimal designs for estimating individual coefficients in polynomial regressiona functional approach. Journal of Statistical Planning and Inference, 118(1), 201–219.
Dette, H., Melas, V. B., Pepelyshev, A. (2010). Optimal designs for estimating the slope of a regression. Statistics, 44(6), 617–628.
Dette, H., Studden, W. J. (1993). Geometry of \(E\)-optimality. Annals of Statistics, 21(1), 416–433.
Dette, H., Wong, W. K. (1996). Robust optimal extrapolation designs. Biometrika, 83(3), 667–680.
Elfving, G. (1952). Optimal allocation in linear regression theory. The Annals of Mathematical Statistics, 23, 255–262.
Fang, Z. (2002). \(D\)-optimal designs for polynomial regression models through origin. Statistics and Probability Letters, 57, 343–351.
Hoel, P. G. (1958). Efficiency problems in polynomial estimation. Annals of Mathematical Statistics, 29(4), 1134–1145.
Hoel, P. G., Levine, A. (1964). Optimal spacing and weighting in polynomial prediction. Annals of Mathematical Statistics, 35(4), 1553–1560.
Huang, M.-N. L., Chang, F.-C., Wong, W. K. (1995). \(D\)-optimal designs for polynomial regression without an intercept. Statistica Sinica, 5(2), 441–458.
Kiefer, J. (1974). General equivalence theory for optimum designs (approximate theory). The Annals of Statistics, 2(5), 849–879.
Mandal, N., Heiligers, B. (1992). Minimax designs for estimating the optimum point in a quadratic response surface. Journal of Statistical Planning and Inference, 31, 235–244.
Melas, V., Pepelyshev, A., Cheng, R. (2003). Designs for estimating an extremal point of quadratic regression models in a hyperball. Metrika, 58, 193–208.
Pronzato, L., Walter, E. (1993). Experimental design for estimating the optimum point in a response surface. Acta Applicandae Mathematicae, 33, 45–68.
Pukelsheim, F. (2006). Optimal design of experiments. Philadelphia: SIAM.
Pukelsheim, F., Studden, W. J. (1993). \(E\)-optimal designs for polynomial regression. Annals of Statistics, 21(1), 402–415.
Sahm, M. (1998). Optimal designs for estimating individual coefficients in polynomial regression. Ph.D. thesis, Fakultät für Mathematik, Ruhr-Universität Bochum, Germany.
Studden, W. J. (1968). Optimal designs on Tchebycheff points. Annals of Mathematical Statistics, 39(5), 1435–1447.
Zen, M.-M., Tsai, M.-H. (2004). Criterion-robust optimal designs for model discrimination and parameter estimation in Fourier regression models. Journal of Statistical Planning and Inference, 124, 475–487.
Acknowledgements
This work has been supported in part by the Collaborative Research Center “Statistical modeling of nonlinear dynamic processes” (SFB 823, Teilprojekt C2) of the German Research Foundation (DFG). The work of Viatcheslav Melas and Petr Shpilev was partly supported by Russian Foundation for Basic Research (Project No. 20-01-00096). The authors thank Martina Stein, who typed parts of this manuscript with considerable technical expertise. The authors are also grateful to two referees and an associate editor for their constructive comments on an earlier version of this paper.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendix: Proof of Lemma 2
Appendix: Proof of Lemma 2
The proof of this result can be found in the PhD thesis of Sahm (1998). As it is difficult to get access to this thesis, we repeat his arguments here for the sake of completeness. We begin with two statements, from which the proof of Lemma 2 will easily follow.
Lemma 4
Let \(Q(x)=\prod _{i=1}^n(x-q_i)\) and \(P(x)=\prod _{i=1}^m(x-p_i)\) denote two polynomials of degree n and m, respectively, where \(m=n-1\) or \(m=n\). Assume that \(P\ne Q\) and set \(q_{n+1}=-\infty \). If
then we have \(R(x) :=(P(x) /Q(x) )'<0,\) whenever the polynomial R(x) is defined.
Proof
In the case \(n=1\), the result is nearly obvious: if \(m=n-1=0\) we have \(R(x) =-Q^{-2}(x) <0\) and for \(m=n=1\)
Now we turn to the case \(n>1\) and assume that the pair (P, Q) form a counterexample of minimal degree. This implies in particular that P(x) and Q(x) cannot have a root in common (otherwise their degree would not be minimal). Furthermore, all roots of P(x) and Q(x) must be simple. As the pair (P, Q) is a counterexample, there exists some \(z\in \mathbb {R}\) where \(R(z)\ge 0.\)
The idea is following. Move the polynomial P up (or down) without changing the property \(R(z)\ge 0\) until one of the zeros of P(x) and Q(x) coincide. Then divide the polynomials by this factor and produce a counterexample of smaller degree, which contradicts the assumption of minimality. For this purpose, we have to consider the two cases \(Q'(z)\ge 0\) and \(Q'(z)< 0\) separately.
We restrict ourselves to the case \(Q'(z)\ge 0\) and mention that the case \(Q'(z)< 0\) can be obtained by exactly the same arguments. To be precise, define
This set is not empty due to the continuity of the roots of the polynomial \( P(x) -\delta \) with respect to \(\delta \). Now let \(\bar{P} (x) =P (x) -\epsilon .\) Then it is clear from the definition of \(\epsilon \) that \(\bar{P}(x) \) has m zeros \(\bar{p}_1, \ldots , \bar{p}_m,\) which interlace with those of Q(x), and \(\bar{P}(x)\) and Q(x) have at least one zero in common. Furthermore, since \(\bar{P}'(x)=P'(x)\) we have
Before we divide \(\bar{P}(x)\) and Q(x) by the factors that they have in common, note that \(\bar{P}\) cannot equal Q because in this case we would have \(P(x)=Q(x)+\epsilon \), which contradicts the interlacing property for \(n>1.\) Now let \(\tilde{P}(x)\) and \(\tilde{Q}(x)\) denote the polynomials obtained by dividing \(\bar{P}(x)\) and Q(x) by their greatest common denominator. These polynomials still have the interlacing property, are of degree smaller than n and are not equal. For the corresponding ratio, we find
which contradicts the assumption that the pair (P, Q) forms a counterexample of minimal degree. \(\square \)
Theorem 7
Consider two polynomials \(Q(x)=\prod _{i=1}^n(x-q_i),\) and \(P(x)=\prod _{i=1}^m(x-p_i),\) of degree n and m, where \(m\le n\le m+1\), the roots \(q_1>\cdots > q_n,\) \(p_1>\cdots > p_m,\) fulfill condition (25) and \(P\ne Q\). Then the zeros of \(q_1'>\cdots >q'_{n-1}\) and \(p_1'>\cdots >p'_{m-1},\) the derivatives \(Q'(x)\) and \(P'(x)\) satisfy
(Here \(q_n'\) id defined as \(q_n'=-\infty \).) In other words, if the polynomials \(P\ne Q\) have only simple roots, which interlace, then the roots of their derivatives strictly interlace.
Proof
We first consider the case where the polynomial P(x) has degree \(m=n-1\). From Lemma 4 for \(i=1,\ldots ,n-1\)
Since the denominator alternates in sign (the roots of \(Q'\) interlace with those of Q) and the leading coefficient is 1, it follows that \(\text {sign}( Q(q_i'))=(-1)^i.\) This implies that the numerator must also alternate in sign, i.e.,
This means that between any pair of consecutive roots \(q_i',\) \(q_{i+1}'\) of \(Q'(x)\) there is a root of \(P'(x)\), which proves Theorem 7 for \(n=m+1.\)
If both polynomials P(x) and Q(x) have the same degree m, note that
Consequently, \(P'(x)\) also has a root in the interval \((-\infty ,q_{m-1}'),\) which completes the proof of Theorem 7. \(\square \)
Proof of Lemma 4
Without loss of generality, we assume that the leading coefficients of the polynomial \(P_1\) and \(P_2\) are equal to 1. With the notation \( P(x)=P_2(x),\) \(Q(x)=P_1(x)\), \(p_i=t_{(n-i+1,2)}, \) \(q_i=t_{(n-i+1,1)} \) (\(i=1\ldots ,n\)) the proposition of Lemma 2 follows from Theorem 7.\(\square \)
About this article
Cite this article
Dette, H., Melas, V.B. & Shpilev, P. Some explicit solutions of c-optimal design problems for polynomial regression with no intercept. Ann Inst Stat Math 73, 61–82 (2021). https://doi.org/10.1007/s10463-019-00736-0
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10463-019-00736-0