1
$\begingroup$

Let $\left[\alpha,\beta\right]$ be a non-empty interval and consider an $\left(n+1\right)$-dimensional subspace $\mathbb{S}_{n}$ of $C^n\left(\left[\alpha,\beta\right]\right)$ such that $\mathbb{S}_{n}$ also has a strictly totally positive normalized basis \begin{align} \mathcal{B}_n:=\left\{b_{n,i}\left(u\right): u \in \left[\alpha, \beta\right]\right\}_{i=0}^{n} \end{align} that is formed by unimodal basis functions, i.e., the row-stochastic matrix \begin{align} B:=\left[b_{n,i}\left(u_j\right)\right]_{j=0,\,i=0}^{n,\,n} &:= \left[ \begin{array}{cccc} b_{n,0}\left(u_0\right) & b_{n,1}\left(u_0\right) & \cdots & b_{n,n}\left(u_0\right) \\ b_{n,0}\left(u_1\right) & b_{n,1}\left(u_1\right) & \cdots & b_{n,n}\left(u_1\right) \\ \vdots & \vdots & \ddots & \vdots \\ b_{n,0}\left(u_n\right) & b_{n,1}\left(u_n\right) & \cdots & b_{n,n}\left(u_n\right) \end{array} \right] \end{align} is strictly totally positive for any sequence \begin{align} \alpha < u_0 < u_1 < \ldots < u_n < \beta \end{align} of strictly increasing knot values $\left\{u_j\right\}_{j=0}^n$ and there exist unique parameter values (maximum points) $\left\{ \overline{u}_i \right\}_{i=0}^n \subset \left[\alpha,\beta\right]$ for which \begin{align} b_{n,i}^{\left( 1\right) }\left( \overline{u}_{i}\right) &=0,\\ b_{n,i}^{\left( 2\right) }\left( \overline{u}_{i}\right) &<0,\\ b_{n,i}^{\left(1\right) }\left( u\right) &>0,~\forall u\in\left[ \alpha,\overline{u}_{i}\right),\\ b_{n,i}^{\left( 1\right) }\left( u\right) &<0,~\forall u\in\left(\overline{u}_{i},\beta\right] \end{align} for all $i=0,1,\ldots,n$.

For example, if one considers the vector space \begin{align} \mathbb{S}_n = \left\langle 1,u,\ldots,u^n : u \in \left[\alpha,\beta\right]\right\rangle \end{align} of polynomials of degree at most $n$, then its unique strictly totally positive normalized basis is formed by the Bernstein polynomials \begin{align} b_{n,i}\left(u\right) = \binom{n}{i} \left(\frac{u - \alpha}{\beta - \alpha}\right)^i \left(1-\frac{u - \alpha}{\beta - \alpha}\right)^{n-i},~u\in\left[\alpha,\beta\right] \end{align} of degree $n$ and $\overline{u}_i = \alpha + \frac{i}{n} \cdot \left(\beta - \alpha\right)$ for all $i=0,1,\ldots,n$.

Note that, in non-polynomial or mixed vector spaces of functions, in general, we do not know the explicit analytical form of the basis functions $\left\{b_{n,i}\right\}_{i=0}^n$.

Consider now a sufficiently smooth non-negative kernel function $\rho\left(u;\sigma\right)$, $u \in \mathbb{R}$ of shape parameter $\sigma \geq 0$ and based on the maximum points $\left\{\overline{u}_i\right\}_{i=0}^{n}$ define the system \begin{align} \mathcal{R}_n^{\sigma}:=\left\{\rho_{n,i}\left(u;\sigma\right):=\rho\left(u-\overline{u}_i; \sigma\right): u \in \left[\alpha,\beta\right]\right\}_{i=0}^n \end{align} of radial basis functions such that:

  • the length of the support $\operatorname{supp}\left(\rho\right)=\overline{\left\{u \in \mathbb{R} : \rho\left(u;\sigma\right) \neq 0\right\}}$ is at least $2\left(\beta - \alpha\right)$;
  • $\rho$ is a bell-shaped unimodal even function;
  • and, concerning the shape parameter $\sigma \geq 0$, one also has the properties: \begin{align*} \rho\left(u; 0\right) &= 1,~ \forall u \in \mathbb{R},\\ \\ \lim_{\sigma \to \infty }\rho\left(u; 0\right) &= \left\{\begin{array}{ll}0,&u \neq 0,\\ 1,&u = 0.\end{array}\right. \end{align*}

For example, one may consider the Gaussian kernel function $\rho\left(u;\sigma\right) = e^{-\left(\sigma u\right)^2}$, $u \in \mathbb{R}$, $\sigma \geq 0$, but my question below is independent of the type of the applied kernel function.


Question: what further conditions (if any) of the (parent) kernel function $\rho$ and of its shape parameter $\sigma$ ensure the strictly total positivity of the perturbed function system \begin{align} \widetilde{\mathcal{B}}_n:=\mathcal{B}_n \circ \mathcal{R}_n^{\sigma}:=\left\{\widetilde{b}_{n,i}\left(u;\sigma\right):= b_{n,i}\left(u\right) \cdot \rho_{n,i}\left(u; \sigma\right) : u \in \left[\alpha,\beta\right] \right\}_{i=0}^{n}, \end{align} i.e., when is strictly totally positive the Hadamard (entry-wise) product of matrices $B$ and \begin{align} R_{\sigma}:=\left[\rho_{n,i}\left(u_j;\sigma\right)\right]_{j=0,\,i=0}^{n,\,n} &:= \left[ \begin{array}{cccc} \rho_{n,0}\left(u_0;\sigma\right) & \rho_{n,1}\left(u_0;\sigma\right) & \cdots & \rho_{n,n}\left(u_0;\sigma\right) \\ \rho_{n,0}\left(u_1;\sigma\right) & \rho_{n,1}\left(u_1;\sigma\right) & \cdots & \rho_{n,n}\left(u_1;\sigma\right) \\ \vdots & \vdots & \ddots & \vdots \\ \rho_{n,0}\left(u_n;\sigma\right) & \rho_{n,1}\left(u_n;\sigma\right) & \cdots & \rho_{n,n}\left(u_n;\sigma\right) \end{array} \right] \end{align} for any strictly increasing knot values $\left\{u_j\right\}_{j=0}^n \subset \left(\alpha,\beta\right)$? Is this even possible at least for a class of kernel functions?


Thank you for your time and energy for dealing with my question.

$\endgroup$

0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.