% !TeX program = pdflatex % !TeX encoding = UTF-8 \documentclass[10pt,a4paper]{article} \usepackage{xurl} \usepackage{singlecolumn} \usepackage{amsthm,mathtools} \clubpenalty=10000 \widowpenalty=10000 \displaywidowpenalty=10000 \numberwithin{equation}{section} \theoremstyle{definition} \newtheorem{theorem}{Theorem}[section] \newtheorem{lemma}[theorem]{Lemma} \newtheorem{proposition}[theorem]{Proposition} \newtheorem{corollary}[theorem]{Corollary} \theoremstyle{remark} \newtheorem{remark}[theorem]{Remark} \newcommand{\R}{\mathbb{R}} \newcommand{\C}{\mathbb{C}} \newcommand{\Q}{\mathbb{Q}} \newcommand{\ii}{\mathrm{i}} \newcommand{\Ree}{\operatorname{Re}} \newcommand{\pair}{\mathcal{P}} \newcommand{\bx}{\mathcal{K}} \title{\vspace{-1.2em}\bfseries All-Order Robust Strictly Positive Real Synthesis for\\Wide-Interval Polynomial Families\vspace{-0.4em}} \author{\normalsize Tiancheng Xu\textsuperscript{1}\quad Guowei Dou\textsuperscript{1}\quad Xiang Ji\textsuperscript{1}\quad ZhengTai Xiong\textsuperscript{1}\quad Long Wang\textsuperscript{2,3}\quad Wensheng Yu\textsuperscript{1}} \date{\vspace{-1.2em}} \renewcommand{\runningtitle}{All-Order Robust SPR Synthesis for Wide-Interval Families} \renewcommand{\reporttype}{} \hypersetup{ pdftitle={All-Order Robust Strictly Positive Real Synthesis for Wide-Interval Polynomial Families}, pdfauthor={Tiancheng Xu, Guowei Dou, Xiang Ji, ZhengTai Xiong, Long Wang, and Wensheng Yu} } \begin{document} \maketitle \footnotetext[1]{School of Electronic Engineering, Beijing University of Posts and Telecommunications, Beijing 100876, China.} \footnotetext[2]{Center for Systems and Control, College of Engineering, Peking University, Beijing 100871, China.} \footnotetext[3]{Center for Multi-Agent Research, Institute for Artificial Intelligence, Peking University, Beijing 100871, China.} \begin{abstract} We establish a uniform all-order existence theorem for robust strictly positive real (SPR) synthesis of Hurwitz-stable interval polynomial families. Specifically, for every order $n \ge 1$, we prove that a common real SPR numerator of exact degree $n$ always exists whenever the lower endpoint of each interior coefficient interval is at most $10^{-5}$ times its upper endpoint, while the constant and leading coefficients vary over arbitrary compact positive intervals. The threshold $10^{-5}$ is independent of both the polynomial order $n$ and the stability margin of the family. Under whole-box Hurwitz stability, such wide interior variation forces geometric separation between successive characteristic frequency scales, decoupling the synthesis constraints into $n$ frequency bands; combining Brouwer's fixed-point theorem in small-margin bands with bandwise strict convexity yields a common exact-degree-$n$ SPR numerator across the entire family. \end{abstract} \keywords{strictly positive real synthesis; interval polynomials; Hurwitz stability; common numerator; scale separation} \section{Introduction}\label{sec:intro} Strictly positive real (SPR) transfer functions play a central role in passivity analysis, absolute stability, adaptive control, and output-error system identification~\cite{DasguptaBhagwat1987,YuWang2008}. For a single Hurwitz-stable denominator $a(s)$, choosing $b(s) = a(s)$ trivially yields a transfer function with strictly positive real part on the imaginary axis. When the denominator varies over an uncertain family, however, robust SPR synthesis requires a \emph{single} fixed real numerator $b(s)$ of prescribed degree such that $\Ree[b(\ii\omega)/a(\ii\omega)] > 0$ holds simultaneously for every admissible denominator $a$ and every real frequency $\omega$. For real interval polynomial families, in which each coefficient varies independently in a closed interval, Kharitonov's theorem~\cite{Kharitonov1978} reduces robust Hurwitz stability to four vertex polynomials. For a \emph{prescribed} numerator, checking the SPR condition across the entire family likewise reduces to these four vertices~\cite{AndersonEtAl1990,WangYu2001}. Synthesis is fundamentally harder because the same numerator must satisfy all four vertex inequalities at all frequencies simultaneously. For low-order Hurwitz-stable interval families ($n \le 4$), Anderson et al.~\cite{AndersonEtAl1990}, Betser and Zeheb~\cite{BetserZeheb1993}, and Yu and Wang~\cite{YuWangFourth2001,WangYu2001} established the existence of a common SPR numerator. For convex combinations of two polynomials (segments), common-numerator synthesis is possible in arbitrary order~\cite{YuWangAckermann2004}. Yet an interval box allows all coefficients to vary independently, so understanding structural regimes that guarantee common-numerator synthesis in arbitrary order remains a fundamental problem. In this paper, we identify a uniform structural regime in which equal-degree robust SPR synthesis succeeds at every order $n \ge 1$ (Theorem~\ref{thm:positive}). Specifically, whenever every interior coefficient interval $[l_j, u_j]$ ($1 \le j < n$) of a Hurwitz-stable interval box satisfies $l_j / u_j \le 10^{-5}$---with arbitrary compact positive intervals for $a_0$ and $a_n$---there always exists a common real SPR numerator of exact degree $n$. At first glance, requiring wide coefficient intervals may seem counterintuitive; however, under the joint hypothesis that the \emph{entire} box is Hurwitz stable, wide interior variation forces adjacent characteristic frequency scales apart geometrically. This separation decouples the positive frequency axis into $n$ bands on each of which a single Kharitonov vertex is extremal, allowing a common numerator to be constructed via Brouwer's fixed-point theorem (for bands with arbitrarily small stability margins) and verified via bandwise strict convexity and quadratic lower bounds. The remainder of the paper is organized as follows. Section~\ref{sec:problem} formulates the interval synthesis problem, recalls the four-vertex reduction, and states Theorem~\ref{thm:positive}. Section~\ref{sec:scale-stationary} establishes the four-coefficient stability inequality, geometric scale separation, and stationary-point alignment in the small-margin regime. Section~\ref{sec:bands-proof} develops the frequency band decomposition, tail decay estimates, and bandwise positivity to complete the proof of Theorem~\ref{thm:positive}. Section~\ref{sec:conclusion} concludes with remarks on sharp thresholds and open questions. \section{Problem Formulation and Main Result}\label{sec:problem} Write polynomial coefficients in ascending order and consider the $n$th-order interval polynomial box \begin{equation}\label{eq:box} \bx = \left\{ a(s) = \sum_{j=0}^{n} a_j s^j : l_j \le a_j \le u_j,\ 0 \le j \le n \right\}, \qquad 0 < l_j \le u_j < \infty. \end{equation} A real polynomial $a \in \R[s]$ is \emph{Hurwitz stable} if all of its complex roots lie in the open left half-plane $\{s \in \C : \Ree s < 0\}$, and the box $\bx$ is \emph{Hurwitz} (or robustly Hurwitz stable) if every member $a \in \bx$ is Hurwitz stable. Following the equal-degree convention in~\cite{YuWang2008}, a real polynomial $b \in \R[s]$ of exact degree $n$ is a \emph{common SPR numerator} for $\bx$ if \begin{equation}\label{eq:target} \Ree\frac{b(\ii\omega)}{a(\ii\omega)} > 0 \qquad (\forall\,a \in \bx,\ \forall\,\omega \in \R). \end{equation} Setting $t = \omega^2 \ge 0$, decompose any real polynomial $a(s) = \sum_{j=0}^n a_j s^j$ into its even and odd parts along the imaginary axis: \begin{equation}\label{eq:parity} a(\ii\omega) = E_a(t) + \ii\omega O_a(t), \qquad E_a(t) = \sum_{k \ge 0} a_{2k}(-t)^k, \qquad O_a(t) = \sum_{k \ge 0} a_{2k+1}(-t)^k. \end{equation} For a candidate numerator $b(s) = \sum_{j=0}^n b_j s^j$, define the real frequency pairing polynomial \begin{equation}\label{eq:row} \pair_{a,b}(t) \coloneqq E_a(t)E_b(t) + t\,O_a(t)O_b(t). \end{equation} Whenever $a$ is Hurwitz stable, $|a(\ii\omega)|^2 = E_a(\omega^2)^2 + \omega^2 O_a(\omega^2)^2 > 0$, so \begin{equation}\label{eq:re-ratio} \Ree\frac{b(\ii\omega)}{a(\ii\omega)} = \frac{\pair_{a,b}(\omega^2)}{|a(\ii\omega)|^2}, \end{equation} and the sign of $\Ree[b(\ii\omega)/a(\ii\omega)]$ coincides with the sign of $\pair_{a,b}(\omega^2)$. \begin{lemma}[Four-vertex reduction~\cite{AndersonEtAl1990,WangYu2001,YuWang2008}]\label{lem:four-vertex} For each fixed $t \ge 0$, as $a$ ranges over $\bx$, the pair $\bigl(E_a(t), O_a(t)\bigr)$ sweeps a closed axis-aligned rectangle whose four corners are realized by four Kharitonov vertex polynomials in $\bx$ independent of $t$. Consequently, for any fixed $b \in \R[s]$ and $t \ge 0$, $\pair_{a,b}(t) > 0$ (respectively, $\ge 0$) holds for all $a \in \bx$ if and only if it holds at these four vertices. \end{lemma} We now state the main theorem of this paper. \begin{theorem}[All-order synthesis under wide interior intervals]\label{thm:positive} Let $n \ge 1$, and let $\bx$ be a Hurwitz box of the form~\eqref{eq:box}. If the interior coefficient intervals satisfy \begin{equation}\label{eq:width} \frac{l_j}{u_j} \le 10^{-5} \qquad (1 \le j < n), \end{equation} then there exists a real polynomial $b \in \R[s]$ of exact degree $n$ satisfying~\eqref{eq:target}. \end{theorem} \begin{remark}[Uniformity and direction of the width condition]\label{rem:uniformity} The threshold~\eqref{eq:width} in Theorem~\ref{thm:positive} is uniform in both the order $n$ and the stability margin of $\bx$, and imposes no width restriction on the constant coefficient $a_0$ or the leading coefficient $a_n$. Note the direction of the inequality:~\eqref{eq:width} requires each interior coefficient to vary across a factor of at least $10^5$. Under the joint hypothesis that the \emph{entire} box is Hurwitz stable, such wide variation forces adjacent characteristic frequency scales apart by a factor of at least $10^{10}$, decoupling the synthesis constraints across frequency bands. \end{remark} \section{Scale Separation and Stationary-Point Alignment}\label{sec:scale-stationary} Since the cases $n = 1$ and $n = 2$ of Theorem~\ref{thm:positive} are elementary (handled at the end of Section~\ref{subsec:global-pos}), we assume throughout Sections~\ref{sec:scale-stationary} and~\ref{subsec:bands} that $n \ge 3$ and that $\bx$ is a Hurwitz box satisfying~\eqref{eq:width}. Define the coefficient ratios, maximum interior ratio, and four-coefficient stability ratios by \begin{equation}\label{eq:xi-D-def} \xi_j \coloneqq \frac{l_j}{u_j} \quad (0 \le j \le n), \qquad \epsilon \coloneqq \max_{1 \le j < n} \xi_j \le 10^{-5}, \qquad D_j \coloneqq \frac{l_j l_{j+1}}{u_{j-1} u_{j+2}} \quad (1 \le j \le n-2). \end{equation} A key structural feature of the proof is the distinction between the \emph{small-margin regime} $1 < D_j \le 4$ (where the stability margin can be arbitrarily small, so crude approximations can lose positivity) and the \emph{large-margin regime} $D_j > 4$ (where direct coefficient dominance suffices). \subsection{Four-Coefficient Ratios and Baseline Scale Separation}\label{subsec:scale-sep} \begin{lemma}[Four-coefficient stability inequality]\label{lem:Dj-gt-one} Every Hurwitz-stable polynomial $a(s) = \sum_{j=0}^n a_j s^j$ satisfies $a_j a_{j+1} > a_{j-1} a_{j+2}$ for all $1 \le j \le n-2$. Consequently, for the Hurwitz box $\bx$, every four-coefficient ratio satisfies $D_j > 1$ ($1 \le j \le n-2$). \end{lemma} \begin{proof} Set $P(x) = E_a(-x)$ and $Q(x) = O_a(-x)$. By Hermite--Biehler strict interlacing~\cite{Holtz2003}, the roots of $P$ and $x Q$ are simple, nonpositive, and strictly alternating, so both $Q/P$ and $P/(xQ)$ admit partial-fraction expansions of the form $c + \sum w/(x+\alpha)$ with $w > 0$ and $\alpha \ge 0$ (in $P/(xQ)$, the residue at $x = 0$ is $P(0)/Q(0) = a_0/a_1 > 0$). For either quotient $A/B$ and each pole $-\alpha$, set $H = B/(x+\alpha)$ (taking out-of-range coefficients to be zero). Comparing coefficients in $A = cB + \sum w H$ yields \[ A_k B_{k+1} - B_k A_{k+1} = \sum w \bigl(H_k^2 - H_{k-1}H_{k+1}\bigr). \] Since all roots of $H$ are real and nonpositive, Newton's inequalities give $H_k^2 - H_{k-1}H_{k+1} \ge 0$, with strict inequality whenever $H_k > 0$. For $Q/P$, every $H$ has positive coefficients throughout its degree range; for $P/(xQ)$, the pole at zero alone contributes $H = Q$. Reading these two sets of inequalities in alternating index order gives $a_j a_{j+1} > a_{j-1} a_{j+2}$ for every $1 \le j \le n-2$. Choosing $a_j = l_j$, $a_{j+1} = l_{j+1}$, $a_{j-1} = u_{j-1}$, and $a_{j+2} = u_{j+2}$ in $\bx$ yields $D_j > 1$. \end{proof} \begin{lemma}[Baseline scales and geometric separation]\label{lem:separation} There exist positive baseline coefficients $\beta_0, \ldots, \beta_n > 0$ satisfying \begin{equation}\label{eq:betaineq} \frac{\beta_j}{\beta_{j+1}} \le \frac{l_{j+1}}{u_{j+2}} \quad (0 \le j \le n-2), \qquad \frac{\beta_{j+1}}{\beta_j} \le \frac{l_j}{u_{j-1}} \quad (1 \le j \le n-1). \end{equation} Writing $W_j \coloneqq u_j / u_{j+2}$ ($0 \le j \le n-2$) and $\tau_j \coloneqq \beta_{j-1} / \beta_{j+1}$ ($1 \le j \le n-1$), we have the geometric scale separation \begin{equation}\label{eq:scales} \frac{W_j}{W_{j+1}} \le \epsilon^2 \quad (0 \le j \le n-3), \qquad \frac{\tau_j}{\tau_{j+1}} \le \epsilon^2 \quad (1 \le j \le n-2). \end{equation} \end{lemma} \begin{proof} Fix $\delta = 1/16$, set $\beta_1 = 1$ and $\beta_0 = \delta l_1 / u_2$, and define the interior ratios inductively by \begin{equation}\label{eq:betas} \frac{\beta_{i+1}}{\beta_i} = \begin{cases} \dfrac{u_{i+2}}{l_{i+1}}, & 1 < D_i \le 4, \\[8pt] \sqrt{\dfrac{u_{i+2} l_i}{l_{i+1} u_{i-1}}}, & D_i > 4, \end{cases} \qquad (1 \le i \le n-2), \end{equation} and finally $\beta_n = \delta l_{n-1}\beta_{n-1}/u_{n-2}$. For each $1 \le i \le n-2$, introduce the dimensionless quantities \begin{equation}\label{eq:AiBi} A_i \coloneqq \frac{u_{i+2}\beta_i}{l_{i+1}\beta_{i+1}}, \qquad B_i \coloneqq \frac{u_{i-1}\beta_{i+1}}{l_i\beta_i}. \end{equation} By~\eqref{eq:betas} and Lemma~\ref{lem:Dj-gt-one}, $(A_i, B_i) = (1, 1/D_i)$ when $1 < D_i \le 4$, and $A_i = B_i = 1/\sqrt{D_i} < 1/2$ when $D_i > 4$. Together with $\beta_0/\beta_1 = \delta l_1/u_2 < l_1/u_2$ and $\beta_n/\beta_{n-1} = \delta l_{n-1}/u_{n-2} < l_{n-1}/u_{n-2}$, this proves~\eqref{eq:betaineq}. Finally, \[ \frac{W_j}{W_{j+1}} = \frac{u_j u_{j+3}}{u_{j+1} u_{j+2}} = \frac{\xi_{j+1}\xi_{j+2}}{D_{j+1}} \le \epsilon^2, \] and multiplying $\beta_{j-1}/\beta_j \le l_j/u_{j+1}$ with $\beta_{j+2}/\beta_{j+1} \le l_{j+1}/u_j$ (from the second inequality in~\eqref{eq:betaineq} at index $j+1$) gives $\tau_j/\tau_{j+1} = (\beta_{j-1}/\beta_j)(\beta_{j+2}/\beta_{j+1}) \le (l_j/u_{j+1})(l_{j+1}/u_j) = \xi_j \xi_{j+1} \le \epsilon^2$, establishing~\eqref{eq:scales}. \end{proof} \subsection{Stationary-Point Alignment in the Small-Margin Regime}\label{subsec:stationary} For each $0 \le i < n$, let $a^{(i)} \in \bx$ denote the Kharitonov vertex taking lower endpoints $l_j$ at indices $j \equiv i, i+1 \pmod{4}$ and upper endpoints $u_j$ at all other indices. As shown in Lemma~\ref{lem:bands} below, $a^{(i)}$ will be the minimizing vertex of $\pair_{a,b}(t)$ on the $i$th frequency band. To handle indices with $1 < D_i \le 4$, we perturb the baseline coefficients $\beta_j$ by positive multipliers $z_j$ ($b_j = \beta_j z_j$) with ratios $x_i = z_{i+1}/z_i$ chosen via Brouwer's fixed-point theorem so that the normalized pairing on band $i$ has a stationary point precisely at a zero of the rotated real part of $a^{(i)}(\ii\omega)$. \begin{proposition}\label{prop:stationary} There exist positive coefficients $b_j = \beta_j z_j$ ($0 \le j \le n$) with $z_0 = z_1 = 1$ and $z_n = z_{n-1}$ such that the interior ratios $x_i \coloneqq z_{i+1}/z_i$ ($1 \le i \le n-2$) satisfy $|x_i - 1| \le 4000\epsilon$, with $x_i = 1$ whenever $D_i > 4$. Moreover, for each $1 \le i \le n-2$ with $D_i \le 4$, setting $T_i \coloneqq l_i / u_{i+2}$, the normalized pairing \begin{equation}\label{eq:normalized} H_i(v) \coloneqq \frac{\pair_{a^{(i)}, b}(T_i v)}{b_i l_i (T_i v)^i} \end{equation} has a stationary point $v_i \in [3/4, 5/4]$ with $H_i(v_i) > 0$. \end{proposition} \begin{proof} Let $x = (x_1, \ldots, x_{n-2})$ range over the cube $[7/8, 9/8]^{n-2}$ and recover $z_0, \ldots, z_n$ from $z_0 = z_1 = 1$, $z_{i+1} = x_i z_i$, and $z_n = z_{n-1}$. Adjacent $z_j$ differ by at most a factor of $8/7$ in either direction. Fix an index $i$ with $1 < D_i \le 4$, and set $t = T_i v = \omega^2$ ($\omega > 0$) for $v \in [1/2, 2]$. Divide both $a^{(i)}(\ii\omega)$ and $b(\ii\omega)$ by $(\ii\omega)^i$, and express their real and imaginary parts in units of $(l_i, \omega l_{i+1})$ for $a^{(i)}$ and $(b_i, \omega \beta_{i+1} z_i)$ for $b$. Using $a^{(i)}_i = l_i$, $a^{(i)}_{i+1} = l_{i+1}$, $a^{(i)}_{i+2} = u_{i+2}$, $a^{(i)}_{i-1} = u_{i-1}$, and $T_i = l_i / u_{i+2}$, these four normalized components take the form \begin{equation}\label{eq:local} R_a(v) = 1 - v + e_1(v), \quad I_a(v) = 1 - (D_i v)^{-1} + e_2(v), \qquad R_b(v) = 1 + e_3(v), \quad I_b(v) = x_i + e_4(v). \end{equation} Define the $C^1$ norm $\|f\|_1 \coloneqq \sup_{v \in [1/2, 2]} |f(v)| + \sup_{v \in [1/2, 2]} |f'(v)|$. We claim that $\|e_1\|_1, \|e_2\|_1 < 64\epsilon^2$ and $\|e_3\|_1, \|e_4\|_1 < 64\epsilon$. Indeed: \begin{itemize}[leftmargin=1.8em,itemsep=2pt] \item In $R_a$, the nearest omitted lower term at $v = 1$ has size at most $u_{i-2}u_{i+2}/l_i^2 = \xi_{i-1}\xi_{i+1}/(D_{i-1}D_i) \le \epsilon^2$, and the nearest upper term has size at most $T_i / W_{i+2} \le \epsilon^2$ (since $T_i = \xi_i W_i$ and $W_i / W_{i+2} \le \epsilon^4$). \item In $I_a$, the nearest omitted lower term at $v = 1$ (present when $i \ge 3$) has magnitude $(\xi_{i-3}/D_i)\,W_{i-3}/T_i$, which is bounded above by $W_{i-3}/T_i \le \epsilon^2$ (using $W_{i-1}/T_i = \xi_{i+1}/D_i$ and~\eqref{eq:scales}), and the nearest omitted upper term has magnitude at most $\xi_i \xi_{i+2}/D_{i+1} \le \epsilon^2$. \item By~\eqref{eq:scales}, each subsequent step outwards in $R_a$ or $I_a$ introduces an additional factor of at most $\epsilon^2$ (out-of-range terms are absent). \item In $R_b$ and $I_b$,~\eqref{eq:betaineq} and the first branch of~\eqref{eq:betas} give $\tau_i / T_i \le \epsilon$ and $T_i / \tau_{i+1} \le \epsilon$. Accounting for the factor $(8/7)^2$ from adjacent $z$-ratios, the nearest omitted terms at $v = 1$ cost at most $3\epsilon$, and each subsequent parity step costs at most $2\epsilon^2$. \end{itemize} Bounding $v^k$ and its derivative on $[1/2, 2]$ yields the geometric majorants \[ \|e_1\|_1, \|e_2\|_1 \le 8\epsilon^2 \sum_{h \ge 0} (5 + 2h)(2\epsilon^2)^h < 64\epsilon^2, \qquad \|e_3\|_1, \|e_4\|_1 \le 12\epsilon \sum_{h \ge 0} (3 + 2h)(4\epsilon^2)^h < 64\epsilon. \] Since $R_a(v) = 1 - v + e_1(v)$ with $\|e_1\|_1 < 64\epsilon^2$, $R_a$ has a unique simple zero $v_i \in [3/4, 5/4]$ with $R_a'(v_i) < 0$; note that $v_i$ depends only on the vertex $a^{(i)}$ and not on $x$. At this zero, $I_a(v_i) > 0$: along the positive imaginary axis, the argument of any Hurwitz polynomial is strictly increasing (each root $-p + \ii q$ with $p > 0$ contributes $p/[p^2 + (\omega - q)^2] > 0$ to $\frac{\mathrm{d}}{\mathrm{d}\omega}\arg a(\ii\omega)$), and dividing by $(\ii\omega)^i$ merely subtracts a constant; hence at any crossing where the real part decreases through zero, the imaginary part must be strictly positive. Using $\beta_{i+1}/\beta_i = u_{i+2}/l_{i+1} = l_i/(l_{i+1} T_i)$ from~\eqref{eq:betas}, the normalized pairing~\eqref{eq:normalized} equals \[ H_i(v) = R_a(v) R_b(v) + v\,I_a(v) I_b(v). \] Its principal part $(1 - v) + (v - D_i^{-1})x_i$ has derivative $x_i - 1$. Using the submultiplicativity of $\|\cdot\|_1$ (with bounds $2, 5, 1, 2$ for the four principal components and $3$ for $v$ on $[1/2, 2]$), we obtain \[ \bigl|H_i'(v_i) - (x_i - 1)\bigr| \le 4000\epsilon < \frac{1}{8}. \] Define the continuous map $F : [7/8, 9/8]^{n-2} \to \R^{n-2}$ by setting $F_i(x) = H_i'(v_i)$ when $D_i \le 4$, and $F_i(x) = x_i - 1$ when $D_i > 4$. Because $|(x_i - F_i(x)) - 1| \le 4000\epsilon < 1/8$, the map $x \mapsto x - F(x)$ maps the compact convex cube $[7/8, 9/8]^{n-2}$ continuously into itself. By Brouwer's fixed-point theorem, there exists a fixed point $x^*$ at which $F(x^*) = 0$. This gives $x_i^* = 1$ for $D_i > 4$, $|x_i^* - 1| \le 4000\epsilon$ for all $i$, and $H_i'(v_i) = 0$ for all $D_i \le 4$. Finally, at $v_i$, $R_a(v_i) = 0$ and $I_b(v_i) \ge 7/8 - 64\epsilon > 0$, so $H_i(v_i) = v_i I_a(v_i) I_b(v_i) > 0$, using the true sign of $I_a(v_i)$ regardless of how small $D_i - 1$ may be. \end{proof} \section{Frequency Band Decomposition and Proof of Theorem~\ref{thm:positive}}\label{sec:bands-proof} \subsection{Frequency Band Decomposition and Envelope Decay}\label{subsec:bands} Fix the numerator $b(s) = \sum_{j=0}^n b_j s^j$ constructed in Proposition~\ref{prop:stationary}. Its alternating parity zeros partition $[0, \infty)$ into $n$ frequency bands on each of which a single vertex $a^{(i)}$ minimizes $\pair_{a,b}(t)$. \begin{lemma}[Frequency band decomposition]\label{lem:bands} There exist strictly increasing nodes $0 = t_0 < t_1 < \cdots < t_{n-1} < t_n = \infty$ such that, for each $0 \le i < n$, the vertex $a^{(i)}$ minimizes $\pair_{a,b}(t)$ over $\bx$ at every finite $t \in [t_i, t_{i+1}]$. Putting $\sigma_j \coloneqq b_{j-1}/b_{j+1}$ ($1 \le j < n$), we have \[ \frac{\sigma_j}{2} < t_j < 2\sigma_j \qquad (1 \le j < n), \] and whenever $1 \le i \le n-2$ satisfies $D_i \le 4$, the stationary point satisfies $T_i v_i \in (t_i, t_{i+1})$. \end{lemma} \begin{proof} Since adjacent $z$-ratios differ by at most $(8/7)^2$,~\eqref{eq:scales} implies $\sigma_j / \sigma_{j+1} \le 2\epsilon^2$. Setting $t = \sigma_j v$ ($v \in [1/2, 2]$), the two terms of degrees $j-1$ and $j+1$ in the corresponding parity component ($E_b$ if $j$ is odd, $O_b$ if $j$ is even), normalized by the degree-$(j-1)$ term, give $1 - v$. At $v = 1$, the nearest omitted term on either side has magnitude at most $(2\epsilon^2)^2$, and each subsequent parity step introduces a factor of at most $2\epsilon^2$. The same geometric series bound as in Proposition~\ref{prop:stationary} keeps the $C^1$ error below $128\epsilon^2$, yielding a unique simple zero $t_j \in (\sigma_j/2, 2\sigma_j)$. Meanwhile, the opposite parity component is dominated by its degree-$j$ term (neighboring terms cost at most $2\epsilon^2$ at $v = 1$). Because these $n-1$ intervals are disjoint, they exhaust all zeros of $E_b$ and $O_b$ on $(0, \infty)$. At $t = 0$, both $E_b(0) = b_0 > 0$ and $O_b(0) = b_1 > 0$. As $t$ crosses $t_1, t_2, \ldots, t_{n-1}$ in turn, the zeros alternate between $E_b$ and $O_b$, starting with $E_b$ at $t_1$. Hence on each band $[t_i, t_{i+1}]$ (for $\omega > 0$), both the real and imaginary parts of $b(\ii\omega)/(\ii\omega)^i$ are nonnegative. Minimizing the pairing $\pair_{a,b}(t)$ in these rotated axes selects the lower endpoints at indices $j \equiv i, i+1 \pmod{4}$ and the upper endpoints elsewhere, which is precisely $a^{(i)}$. Continuity includes the band endpoints and $t = 0$. Finally, for $D_i \le 4$, the bounds $\tau_i/T_i, T_i/\tau_{i+1} \le \epsilon$ from Proposition~\ref{prop:stationary} give $\sigma_i/T_i, T_i/\sigma_{i+1} \le 2\epsilon$, placing $T_i v_i \in [\frac{3}{4}T_i, \frac{5}{4}T_i]$ strictly inside $(t_i, t_{i+1})$. \end{proof} For any fixed vertex $a \in \bx$, expand its pairing with $b$ as $\pair_{a,b}(t) = \sum_{k=0}^n c_k t^k$. \begin{lemma}[Coefficient bounds and tail decay]\label{lem:tails} For every vertex $a$ of $\bx$ and every $0 \le k \le n$: \begin{enumerate}[label=(\roman*),leftmargin=2.2em,itemsep=2pt] \item $|c_k| \le 4 b_k u_k$, and if $a_k = u_k$, then $c_k \ge (1 - 4\epsilon) b_k u_k$; \item on band $i$ ($t \in [t_i, t_{i+1}]$), successive upper envelopes $b_j u_j t^j$ decrease by a factor of at most $q_0 \coloneqq 8\epsilon^3$ when moving upwards from $j \ge i+2$, or downwards from $j \le i-1$. \end{enumerate} \end{lemma} \begin{proof} Expanding~\eqref{eq:row} yields \[ c_k = \sum_{h} (-1)^h b_{k-h} a_{k+h}, \] summed over all integers $h$ for which both $0 \le k-h \le n$ and $0 \le k+h \le n$. By~\eqref{eq:betaineq} and $z_{k \pm 1}/z_k \le 8/7$, each distance-one term ($h = \pm 1$) has magnitude at most $(8/7) b_k l_k$. Putting $r_j \coloneqq u_{j+1}/u_j$, we have $r_{p+1}/r_{p-1} = \xi_p \xi_{p+1} / D_p \le \epsilon^2$, so $r_{k+h}/r_{k-h} \le \epsilon^{2h}$ for $h \ge 1$. Hence for $h \ge 1$, the ratio of successive envelopes $b_{k-h} u_{k+h}$ satisfies \[ \frac{b_{k-h-1} u_{k+h+1}}{b_{k-h} u_{k+h}} \le \frac{8}{7}\,\xi_{k-h}\,\frac{r_{k+h}}{r_{k-h}} \le 2\epsilon^{2h}. \] Reversing coefficient order gives the same bound for $h \le -1$. Summing over $|h| \ge 2$, all terms at distance at least two together cost at most $8\epsilon^2 b_k l_k$. Combining the central term $b_k a_k$ with the distance-one and tail bounds yields (i) (with $c_0 = b_0 a_0$ and $c_n = b_n a_n$ exact). For (ii), if a high tail exists on band $i$ ($i \le n-2$), Lemma~\ref{lem:bands} gives $t \le t_{i+1} < 2 b_i / b_{i+2}$. For any $j \ge i+2$, applying~\eqref{eq:betaineq} and $W_{j-1} \ge W_{i+1}$ (from~\eqref{eq:scales}) yields \[ t\,\frac{b_{j+1} u_{j+1}}{b_j u_j} \le 2\left(\frac{8}{7}\right)^3 \frac{l_{i+1} l_{i+2}}{u_{i+2} u_{i+3}}\,\xi_j\,\frac{u_{j+1}}{u_{j-1}} \le 8\epsilon^3 = q_0. \] The low-tail estimate ($j \le i-1$) follows symmetrically by reversing all coefficient indices and replacing $t$ with $1/t$. \end{proof} \subsection{Bandwise Positivity and Proof of Theorem~\ref{thm:positive}}\label{subsec:global-pos} We verify strict positivity of $\pair_{a^{(i)}, b}(t)$ separately on the interior bands and on the two boundary bands, and then assemble the proof of Theorem~\ref{thm:positive}. \begin{lemma}[Positivity on interior bands]\label{lem:interior-bands} For every interior index $1 \le i \le n-2$ and every $t \in [t_i, t_{i+1}]$, we have $\pair_{a^{(i)}, b}(t) > 0$. \end{lemma} \begin{proof} Let $a = a^{(i)}$ and write $\pair_{a,b}(t) = \sum_{k=0}^n c_k t^k$. By definition of $a^{(i)}$, we have $a_{i-1} = u_{i-1}$ and $a_{i+2} = u_{i+2}$, so Lemma~\ref{lem:tails}(i) gives $c_{i-1} \ge (1 - 4\epsilon) b_{i-1} u_{i-1} > 0$ and $c_{i+2} \ge (1 - 4\epsilon) b_{i+2} u_{i+2} > 0$. In the scaled function $g_i(t) \coloneqq t^{-i}\pair_{a,b}(t)$, these two terms carry powers $t^{-1}$ and $t^2$, whose second derivatives are strictly positive, while the central terms $c_i$ and $c_{i+1} t$ vanish upon differentiating twice. By Lemma~\ref{lem:tails}(i)--(ii), relative to the nearer of the two positive second derivatives ($2 c_{i-1} t^{-3}$ on the low side and $2 c_{i+2}$ on the high side), the sum of the absolute values of all more distant second derivatives on the same side is bounded by \begin{equation}\label{eq:tailbudget} \frac{4}{1 - 4\epsilon} \sum_{h \ge 1} \frac{(h+1)(h+2)}{2}\,q_0^h = \frac{4}{1 - 4\epsilon}\bigl[(1 - q_0)^{-3} - 1\bigr] < \frac{1}{2}. \end{equation} Hence $g_i''(t) > 0$ throughout $(t_i, t_{i+1})$, so $g_i$ is strictly convex on $[t_i, t_{i+1}]$. \begin{itemize}[leftmargin=1.8em,itemsep=2pt] \item If $1 < D_i \le 4$, then $g_i(T_i v) = b_i l_i H_i(v)$ attains a strictly positive stationary value at $T_i v_i \in (t_i, t_{i+1})$ by Proposition~\ref{prop:stationary} and Lemma~\ref{lem:bands}. Strict convexity implies $g_i(t) > 0$ on all of $[t_i, t_{i+1}]$. \item If $D_i > 4$, then $x_i = 1$ and $A_i = B_i = 1/\sqrt{D_i} < 1/2$. In $c_i$, the neighbor $b_{i+1} u_{i-1}$ costs $B_i < 1/2$ relative to the central term $b_i l_i$, while the other neighbor $b_{i-1} l_{i+1}$ uses the lower endpoint $l_{i+1}$ and costs at most $(8/7)\epsilon$. Symmetrically, in $c_{i+1}$, the neighbor $b_i u_{i+2}$ costs $A_i < 1/2$ relative to $b_{i+1} l_{i+1}$, and the other neighbor costs at most $(8/7)\epsilon$. Combining with the distance-$\ge 2$ bound $8\epsilon^2$ from Lemma~\ref{lem:tails}, both central coefficients satisfy \[ \frac{c_i}{b_i l_i},\ \frac{c_{i+1}}{b_{i+1} l_{i+1}} \ge 1 - \frac{1}{2} - \frac{8}{7}\epsilon - 8\epsilon^2 > 0. \] Meanwhile, the estimate~\eqref{eq:tailbudget} (without the derivative factors) shows that the remaining tails ($k \le i-2$ and $k \ge i+3$) are bounded in magnitude by less than half of the two positive outer terms $c_{i-1}t^{i-1}$ and $c_{i+2}t^{i+2}$. Therefore $\pair_{a,b}(t) > 0$ on $[t_i, t_{i+1}]$. \qedhere \end{itemize} \end{proof} \begin{lemma}[Positivity on the first and last bands]\label{lem:boundary-bands} We have $\pair_{a^{(0)}, b}(t) > 0$ for all $t \in [0, t_1]$ and $\pair_{a^{(n-1)}, b}(t) > 0$ for all $t \in [t_{n-1}, \infty)$. \end{lemma} \begin{proof} On band $0$, $a = a^{(0)}$ has $a_0 = l_0$, $a_1 = l_1$, and $a_2 = u_2$. The constant and linear coefficients of $\pair_{a,b}(t)$ are $c_0 = b_0 l_0 > 0$ and $c_1 = S - b_0 u_2 - M$, where $S \coloneqq b_1 l_1$ and $M \coloneqq b_2 l_0$. Subtracting the high tail ($k \ge 3$) from $c_2 t^2$ via Lemma~\ref{lem:tails} leaves at least $\frac{1}{2} b_2 u_2 t^2$ (since $1 - 4\epsilon - 4q_0/(1 - q_0) > 1/2$). By construction of $\beta_0$, $b_0 u_2 = \delta S$ with $\delta = 1/16$, while~\eqref{eq:betaineq} and Proposition~\ref{prop:stationary} give $M/S \le 1 + 8000\epsilon$. If $M \le (1 - 2\delta)S$, then $c_1 > 0$ and the quadratic $c_0 + c_1 t + \frac{1}{2} b_2 u_2 t^2$ is strictly positive on $[0, \infty)$. Otherwise $M > (1 - 2\delta)S$, and the negative part of $c_1$ satisfies $(-c_1)_+ \le (\delta + 8000\epsilon)S$; dividing its square by $c_0 \cdot (\frac{1}{2} b_2 u_2) = \frac{1}{2} b_0 u_2 M > \frac{1}{2}\delta(1 - 2\delta)S^2$ yields \[ \frac{(-c_1)_+^2}{\frac{1}{2} c_0 b_2 u_2} < \frac{2(\delta + 8000\epsilon)^2}{\delta(1 - 2\delta)} < 4. \] Hence if $c_1 \ge 0$, the quadratic $c_0 + c_1 t + \frac{1}{2} b_2 u_2 t^2$ is strictly positive on $[0, \infty)$ directly; if $c_1 < 0$, then $(-c_1)_+^2 = c_1^2$, so the bound above shows that its discriminant is strictly negative. In either case, $\pair_{a^{(0)}, b}(t) > 0$ for all $t \in [0, t_1]$. On the last band $i = n-1$, reversing the coefficient order and substituting $t \mapsto 1/t$ yields the exact same quadratic estimate and sign-case split, since $b_n u_{n-2} = \delta b_{n-1} l_{n-1}$ and the reversed multiplier ratio is bounded by $1/(1 - 4000\epsilon) \le 1 + 8000\epsilon$. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:positive}] For $n = 1$, $b(s) = 1 + s$ gives $\pair_{a,b}(t) = a_0 + a_1 t > 0$ for all $t \ge 0$. For $n = 2$, pick $b_1 = 1$ and choose $b_0, b_2 > 0$ small enough that $b_0 u_2 + b_2 u_0 < l_1$; then $\pair_{a,b}(t) = b_2 a_2 t^2 + (a_1 - b_0 a_2 - b_2 a_0)t + b_0 a_0$ has all three coefficients strictly positive across $\bx$. For $n \ge 3$, take the degree-$n$ polynomial $b$ constructed in Proposition~\ref{prop:stationary} (whose leading coefficient $b_n > 0$). By Lemma~\ref{lem:bands}, the bands $[t_i, t_{i+1}]$ ($0 \le i < n$) cover $[0, \infty)$ with minimizing vertex $a^{(i)}$. Lemmas~\ref{lem:interior-bands} and~\ref{lem:boundary-bands} show that $\pair_{a^{(i)}, b}(t) > 0$ on each band, so $\pair_{a,b}(t) > 0$ for all $a \in \bx$ and $t \ge 0$ by Lemma~\ref{lem:four-vertex}. Equation~\eqref{eq:re-ratio} and conjugate symmetry then yield~\eqref{eq:target} for every $a \in \bx$ and every $\omega \in \R$. \end{proof} \section{Conclusion}\label{sec:conclusion} We have established a uniform all-order sufficient condition for equal-degree robust SPR synthesis of Hurwitz-stable interval polynomial families: whenever every interior coefficient interval satisfies $l_j/u_j \le 10^{-5}$ ($1 \le j < n$), whole-box Hurwitz stability forces geometric scale separation across frequency bands and guarantees the existence of a common exact-degree-$n$ SPR numerator in every order $n \ge 1$. Determining the sharp threshold for the interior endpoint ratio $l_j/u_j$ and relaxing the uniform requirement across all interior indices remain natural directions for future work. \section*{Acknowledgements} GPT-6 Astra was used for calculations, proof development, and editorial assistance. AI-generated suggestions were verified against the mathematical arguments and the accompanying formal proof; responsibility for the final manuscript rests with the authors. The main theorem of this paper has been formally verified in Lean~4 (declaration \path{SPR.General.Direct.uniformWideSynthesis} in \path{Lean/Main.lean}), and the source code repository is available at \url{https://github.com/dgjszy/spr-counterexamples}. \bibliographystyle{plain} \bibliography{references} \end{document}