Basis transforms preserving real-rootedness and PF
A basis transform sends $x^n$ to a fixed polynomial $B_n$. For Newton bases, for bases defined by order-two recurrences, and for the semigroup generated by $aD+bxD^2$, we determine exactly when the transform preserves real-rootedness or the Pólya frequency property. In each case a short list of parameter inequalities decides.
Results
For $B_n\in\mathbb R[x]$ let $T_B$ be the linear map with $T_B(x^n)=B_n$, that is, $T_B\bigl(\sum_n a_nx^n\bigr)=\sum_n a_nB_n$. A polynomial is PF (Pólya frequency) if it is zero or has nonnegative coefficients and only real, nonpositive roots. On this page a real-rooted polynomial may be zero.
Theorem (Newton bases). For $c=(c_j)_{j\ge0}$ let $N_n^{c}(x)=\prod_{j=0}^{n-1}(x+c_j)$ and $T_c(x^n)=N_n^{c}$. The following are equivalent:
- $T_c$ maps every real-rooted polynomial with only nonpositive roots to a real-rooted polynomial;
- $T_c\bigl((x+r)^n\bigr)$ is real-rooted for all $n\ge0$ and $r\ge0$;
- $c_j=c_0+jh$ for all $j$, for some $h\ge0$.
For $N\ge2$ the same equivalences hold for inputs of degree at most $N$, with the tests restricted to $n\le N$ and the progression condition to $j<N$.
Theorem (order-two bases). Let $a\ge0$, $b>0$, $s\ge0$ and $c\in\mathbb R$. Put $B_0=1$, $B_1=s$ and $B_{n+2}=a\,B_{n+1}+(bx+c)\,B_n$. The following are equivalent:
- $T_B$ maps every PF polynomial to a PF polynomial;
- every $B_n$ is PF;
- $B_n(0)\ge0$ for all $n$;
- $a^2+4c\ge0$ and ($2s\ge a$ or $c\ge s^2-as$).
The same criterion characterizes the maps that send PF polynomials with roots in $[-R,0]$ to PF polynomials, for any fixed $R\ge0$. No degree bound suffices: for every $N$ there is $c<-\tfrac14$ such that, for $a=b=s=1$, $T_B$ maps PF polynomials of degree at most $N$ to PF polynomials but not every PF polynomial.
Theorem (ray semigroup). For $a,b,t\in\mathbb R$ let $T_t$ be the linear map with $T_t(x^n)=B_n(t)$, where $B_0=1$ and $$ B_{n+1}=\bigl(x+t(a+bn)\bigr)B_n+bt\,x\,B_n'. $$ Then $T_0=\mathrm{id}$ and $T_{s+t}=T_sT_t$, and $T_t$ maps PF polynomials to PF polynomials for every $t\ge0$ if and only if $a\ge0$ and $b\ge0$. The two parameters are read off from the truncated logarithm $\ell=(T_1-1)-\tfrac12(T_1-1)^2$ as $a=[x^0]\,\ell(x)$ and $b=\tfrac12[x^1]\,\ell(x^2)-[x^0]\,\ell(x)$.
The other results are as follows.
- Ray semigroup. For all real $a,b,t$, applying the falling factorial transform of step $bt$ to $T_tp$ gives $(R_{bt}p)(x+ta)$, where $R_\mu(x^n)=\prod_{i<n}(x+i\mu)$. Finite words in translations $p(x)\mapsto p(x+a)$ and Lah flows ($T_1$ with $a=0$ and time $b$), with $a,b\ge0$, and their coefficientwise limits preserve PF polynomials. The monomial images of a word have exponential generating function $\exp\bigl(\sum_{(u,a)}\frac{az}{1-uz}\bigr)\exp\bigl(\frac{xz}{1-\beta z}\bigr)$, where $\beta$ is the total Lah time and each translation $a$ is recorded with the Lah time $u$ applied before it; every datum with $\beta\ge0$, $0\le u\le\beta$, $a\ge0$ and $u$ weakly decreasing arises from a word. If $a,b,\kappa\ge0$, $c_n\ge0$, $S_0=1$ and $S_{n+1}=(x+c_n)S_n+\kappa xS_n'$, then every $T_1S_n$ is PF.
- Interval and ray operators. Let $L<U$, $\rho,\kappa>0$, $\alpha,\beta\in\mathbb R$ and $\mathcal Df=\rho\bigl((\alpha x+\beta)f+\kappa(x-L)(x-U)f'\bigr)$, and let $g$ have positive leading coefficient, all roots in $[L,U]$, and $\alpha+\kappa\deg g>0$. Then $x^n\mapsto\mathcal D^ng$ maps every nonzero real-rooted polynomial with all roots in $S$ to a nonzero real-rooted polynomial with all roots in $[L,U]$ if and only if $$ S\subseteq\bigl[\rho(\alpha L+\beta-\kappa(U-L)m_L),\ \rho(\alpha U+\beta+\kappa(U-L)m_U)\bigr], $$ where $m_L,m_U$ are the multiplicities of $L,U$ as roots of $g$. For $\mathcal Df=(a+bx)f+\kappa(x-h)f'$ with $b>0$, $\kappa\ge0$, and $g\ne0$ real-rooted with roots in $(-\infty,h]$, the target $(-\infty,h]$ is reached exactly when $S\subseteq(-\infty,a+bh+\kappa m_h]$.
- Appell coefficients. Let $P_n$ be monic of degree $n$ such that $T_P$ maps real-rooted polynomials with nonpositive roots to real-rooted polynomials, and write $T_Pp=\sum_kA_kp^{(k)}$. Then $\deg A_k\le\lfloor k/2\rfloor$ and $0\le[x^j]A_{2j}\le\bigl([x^1]A_2\bigr)^j/j!$. The conditions $[x^1]A_2=0$, that every $A_k$ is constant, that $T_P$ commutes with translations, and that $P_n'=nP_{n-1}$ for all $n$ are equivalent, and they imply that $T_P$ maps every real-rooted polynomial to a real-rooted polynomial.
- Eventually geometric bases. Let $P_n=P_ky^{n-k}$ for $n\ge k\ge1$. If $y^kP_j=P_ky^j$ for all $j<k$, then $T_P(p)=P_0\,p(y)$. Otherwise, if $y=a+dx$ with $d\ne0$ and $P_k\ne0$, then for every $r\ne0$ some $N$ in an explicit range makes $T_P\bigl((x+r)^N\bigr)$ not real-rooted.
- Scalar-index recurrences. If $P_0=g$ and $P_{n+1}=\Phi P_n+nh\,P_n$ for a linear map $\Phi$, then $x^n\mapsto P_n$ is the map $x^n\mapsto\prod_{i<n}(x+ih)$ followed by $x^n\mapsto\Phi^ng$.
- Finite symbols. If the finite Borcea–Brändén symbol $\sum_{k\le d}\binom dkP_k(x)\,w^{d-k}$ is stable, then $T_P$ maps real-rooted polynomials of degree at most $d$ to real-rooted polynomials, and PF polynomials to PF polynomials when $P_0,\dotsc,P_d$ have nonnegative coefficients.
Use
The Newton and order-two classifications reduce preservation to tests on repeated roots and constant terms. The multiplier sequences page treats the classical diagonal preservers, and the Narayana polynomials page uses rising and falling factorial bases.
References
- Borcea–Brändén symbols on symmetricfunctions.com
- Multiplier sequences on symmetricfunctions.com