\subsection{Series estimates for \texorpdfstring{$1<\lambda<2/\sqrt3$}{1 < lambda < 2/sqrt(3)}} \label{subsec:low-computer-region} Fix $1<\lambda<\min(m,2/\sqrt3)$. Lemma~\ref{lem:calibration} gives $D_\lambda>0$ when $\alpha=1$. For $0<\alpha<1$, put \[\begin{gathered} q=\frac{u}{1-u}=\frac hc,\qquad R_* =\frac vc,\qquad w_i=\lambda^{i-1}\beta_i^2,\\ T_p=\sum_{i\ge1}w_i,\qquad Z_p=2q-(\lambda+1)T_p,\qquad R=\frac{T_p}{Z_p},\quad \theta=\frac Rq. \end{gathered}\] The denominator $Z_p$ is positive by Lemma~\ref{lem:harmonic-speed}, which also gives $R_*i}w_sW_s(R), \nonumber\\ F_{\rm pipe}&=-2RuL_1+u^2W_1(R)+V_1,\nonumber\\ \tau_i&=1-C_\alpha r_i+(\lambda+1)R-R/r_{i-1},\nonumber\\ B_i&=w_i\left\{\frac{2i(\alpha_p+cR)}\lambda +2iRr_iL_i+2(i+1)R\lambda r_i^2L_i +k(2i-1)\tau_i\right\}, \nonumber\\ S_i&=B_i-2V_i.\label{eq:low-B-S} \end{align} Set $W_1^Y=0$, $W_i^Y=k w_i(2i-1)R/r_{i-1}$ for $i\ge2$, and \begin{equation}\label{eq:low-J} J_i=V_i+k w_i(2i-1)C_\alpha r_i -\frac{W_i^Y+W_{i+1}^Y}{2},\qquad i\ge2. \end{equation} Here $J_i$ is the coefficient of $\Delta_i^2$ in the bound below. The polynomial $J_{s-1}(w)$ in~\eqref{eq:assigned-J} carries the argument $w$. \begin{lemma}[A bound in terms of the differences $\Delta_i$]\label{lem:calibrated-head} Suppose $0<\alpha<1$. Then $V_i>0$ for every $i\ge1$, $\tau_1=0$, and $0<\tau_i<1$ for $i\ge2$. Put \[ \kappa=\frac{(1-u)T_p}{u^2}, \] \[\begin{split} C_{\rm head}={}&2A_0u\{-k\log(a_2/u)-1+a_2/u\}\\ &+\frac{(u-\lambda a_2) \{2-(3\lambda+1)u+\lambda(\lambda+1)a_2\}}\lambda. \end{split}\] Then $C_{\rm head}>0$, $\kappa>1$, and \begin{align} F_{\rm floor}(v)>{}&F_{\rm pipe} +\sum_{i\ge1}\Delta_i\rho_i S_i +e^2\left\{V_1+\frac{2u^2 C_\alpha}{\lambda(1-e)} -\frac{W_2^Y}{2}\right\} +\sum_{i\ge2}\Delta_i^2J_i+C_{\rm cal}e, \label{eq:low-master}\\ C_{\rm cal}&=\frac{C_{\rm head}u^2(\kappa-1)(2-e)}{Z_p}>0. \nonumber \end{align} The coefficient in braces in~\eqref{eq:low-master} is positive. Also \begin{equation}\label{eq:low-e-caps} \begin{gathered} 00 \end{aligned}\] for $n\ge1$. The second function is strictly decreasing on $(0,1/\lambda)$; all its weighted expectations are finite. The derivative is also bounded above by $(A_0/a_2)a_{n+2}$. Thus the difference $\widetilde F(cR)-F_{\rm floor}(cR)$ in~\eqref{eq:strict-Mellin-gap} increases with $R$, and the geometric bounds in Lemma~\ref{lem:mellin-floor} justify termwise differentiation. The first two terms in $-c\widetilde F'(w)$ have the lower bounds \begin{align*} 2c^2\E[g\cJ(1-y)]&\ge2A_0\gamma_0,\\ c\E[b\cJ((1-y)(2-y))]&\ge \frac{(u-\lambda a_2) \{2-(3\lambda+1)u+\lambda(\lambda+1)a_2\}}\lambda. \end{align*} They follow by the $U,V$ substitution and $(1-\lambda UV)^{-2}\ge1$; the remaining factors are positive. Jensen's inequality under the probability weighted by $U/u$ gives $\gamma_0\ge u\{-k\log(a_2/u)-1+a_2/u\}$. The expression in braces is positive: the function $-k\log z-1+z$ vanishes at $z=1/\lambda$ and has negative derivative for $0F_{\rm floor}(cR)+C_{\rm head}(R-R_*). \end{equation} We next bound $R-R_*$ for the same offspring law. The identity $\lambda^{s-1}\beta_s\beta_{s-1}=q(\beta_{s-1}-\beta_s)$, finite summation by parts, and the vanishing last endpoint give \begin{align*} q-Q&=q\sum_{i\ge1}\Delta_i(a_i+a_{i+1}),\\ T_p-T&=\sum_{i\ge1}\Delta_i(\rho_i+\rho_{i+1}) \sum_{s>i}w_s,\\ R-R_*&=\frac{2q}{Z_p Z}\sum_{i\ge1}\Delta_i \left\{(\rho_i+\rho_{i+1})\sum_{s>i}w_s-T_p(a_i+a_{i+1})\right\}. \end{align*} All tail terms in the first two formulas are nonnegative. The ratios $r_i$ increase strictly. With $f_i=(\sum_{s>i}w_s)/\beta_i$, the same telescoping sum gives \[ f_i-f_{i+1}=\lambda^i \left\{\frac{\beta_{i+1}^2}{\beta_i} -\frac1q\sum_{s>i+1}w_s\right\}<0. \] Indeed, each term of the last tail is strictly larger than $r_i$ times the corresponding term $\lambda^{s-1}\beta_s\beta_{s-1}$. Summing the latter over $s>i+1$ gives $q\beta_{i+1}$. Therefore $f_i\ge f_1=T_p/u-u=T_p+u(\kappa-1)$. Solving $\alpha_p<1$ in terms of $\kappa$ gives $\kappa>2\lambda/(\lambda+1)>1$. It follows that every expression in braces in the speed-difference formula is strictly larger than $u(\kappa-1)\beta_i(\rho_i+\rho_{i+1})$. Since $0\frac{u^2(\kappa-1)e(2-e)}{Z_p}. \end{equation} To identify the coefficients of $\Delta_i^2$, put $\bar\alpha=a(cR)$. The definitions of $a(cR)$ and $\alpha_p$ give \[ \bar\alpha-\alpha_p=\frac{C_\alpha e}{1-e}. \] For $n_i=a_{i+1}/a_i$ define $c_i=2i/\lambda-k(2i-1)n_i>1/\lambda$. Substitute $a_i=\beta_i\rho_i$ and $\delta_i=\beta_{i+1}\Delta_i$ in the series \eqref{eq:complete-mellin-floor}. Use $(1+L_i)r_i=1/\lambda$ and sum the diagonal terms by parts. The result, with every end term displayed, is \begin{align} F_{\rm floor}(cR)={}&F_{\rm pipe} +\sum_{i\ge1}\Delta_i\rho_i S_i+\sum_{i\ge1}\Delta_i^2V_i +(\bar\alpha-\alpha_p)\sum_{i\ge1}w_i\rho_i\Delta_i c_i \nonumber\\[-2pt] &+k\sum_{i\ge1}w_i(2i-1) \left\{C_\alpha r_i\Delta_i^2 -\frac{R\Delta_i\Delta_{i-1}}{r_{i-1}}\right\}, \qquad\Delta_0=0.\label{eq:low-square-identity} \end{align} One may check the local expansion before summation by noting that its coefficient of $\Delta_i$ differs from $\rho_i B_i$ by \[ w_i\rho_i(\bar\alpha-\alpha_p)c_i +k w_i(2i-1)\left\{C_\alpha r_i\Delta_i -R\Delta_{i-1}/r_{i-1}\right\}. \] More explicitly, before replacing the diagonal terms by the $V_i$ tails, the difference between the $i$th summand of $F_{\rm floor}(cR)$ and these local terms is \begin{align*} &-(i-1)(1+cR)w_i\rho_i^2 +i(1+cR)w_{i+1}\rho_{i+1}^2\\ &\hspace{5mm} -\frac{2iRw_iL_i}{r_{i-1}}\rho_{i-1}\rho_i +2(i+1)Rw_{i+1}L_i\rho_i\rho_{i+1}. \end{align*} The first pair telescopes to zero. The second pair telescopes because $L_{i+1}/r_i=L_i$, leaving exactly $-2Rw_1L_1/r_0=-2RuL_1$. Its upper endpoint tends to zero geometrically. Finally the diagonal telescope uses $1-\rho_s^2=\sum_{i0$ is proved in the next lemma. Also $C_\alpha/(\lambda R)>1$ and $k\lambda r_1<1$, so \[ \frac{2u^2C_\alpha}{\lambda(1-e)}>2u^2R, \qquad \frac{W_2^Y}{2}<\frac32u^2R. \] This proves positivity of the coefficient in braces in~\eqref{eq:low-master}. Under the distribution $\PP(U_p=\lambda^{-j})=hd^{j-1}$, the quadratic $Q_{\alpha_p}$ has zero mean and changes sign once from negative to positive. The proof of Lemma~\ref{lem:mellin-floor} applied to this scalar distribution gives $\tau_1=0$ and $\tau_i>0$ for $i\ge2$. Furthermore \[ (\lambda+1)-\lambda r_i-1/r_{i-1} =-\frac{(\lambda+1)L_i(c+\lambda L_i)} {(1+L_i)(1-\lambda L_i)}<0, \] so $\tau_i<1$. Substituting $\bar\alpha-\alpha_p=C_\alpha e/(1-e)$ into $\bar\alpha<1$ gives $e0$ gives the stated bound on $\sqrt e$. Absolute convergence follows from $\beta_i,a_i\le\lambda^{-i}$ and the polynomial factors in $i$; finite summation followed by geometric domination justifies all signed rearrangements in this proof. \end{proof} \subsubsection{\texorpdfstring{Bounds for the tails $V_i$ and $\sum_{s>i}w_s$}{Bounds for V(i) and the tail sum of w(s)}} We record the estimates that turn the infinite sums into finite rational expressions. Let $\delta=\log\lambda$ and $a=c/\delta$. Taylor's formula with integral remainder applied to $a=\int_0^1(1+c)^s\dd s$ gives \begin{equation}\label{eq:low-log-enclosure} a_L=1+c/2-c^2/12\le a\le a_U=1+c/2-c^2/12+c^3/24\le1+c/2. \end{equation} For every $z>0$, \begin{align} 1+z/2+z^2/12-z^4/720 &\le\frac{z}{1-e^{-z}}\le1+z/2+z^2/12,\label{eq:low-global-F}\\ 1-z^2/12&\le\frac{z^2}{2(\cosh z-1)} \le1-z^2/12+z^4/240.\label{eq:low-global-C} \end{align} For example, multiply the two inequalities in~\eqref{eq:low-global-F} by the positive exponential denominator. Beyond the vanishing low powers, their coefficients, with the indicated sign, are respectively $(n-3)(n-5)(n-6)(n+8)/(720n!)$ for $n\ge7$ and $(n-3)(n-4)/(12n!)$ for $n\ge5$. For~\eqref{eq:low-global-C}, the corresponding even coefficients are $\{n(n-1)-12\}/(6n!)$ for $n\ge6$ and $(n-6)(n^3-9n-40)/(120n!)$ for $n\ge8$. All omitted lower coefficients vanish. Absolute convergence of the exponential series proves the inequalities on the entire positive axis. \begin{lemma}[Bounds for the tails of the Lambert series]\label{lem:low-Lambert-tails} For every $i\ge1$, $V_i>0$. If $b=\lambda L_i$, put \[ r=\frac1{\lambda+b}=r_i,\qquad A=c+br,\quad B=br,\quad t=\frac r{b+c}, \] and define the five polynomials \begin{align*} P_1&=A,\qquad P_2=A(A+B),\qquad P_3=A(A^2+4AB+B^2),\\ P_4&=A(A^3+11A^2B+11AB^2+B^3),\\ P_5&=A(A^4+26A^3B+66A^2B^2+26AB^3+B^4). \end{align*} Define the following explicit rational functions of an argument $a>0$: \begin{align} T_L(a)&=t\{a+A/2+P_2/(12a)-P_4/(720a^3)\},\nonumber\\ T_U(a)&=t\{a+A/2+P_2/(12a)\},\nonumber\\ P_U(a)&=\frac t\lambda \{(i+1)(aA+P_2/2+P_3/(12a))+a^2-P_2/12+P_4/(240a^2)\},\nonumber\\ P_L(a)&=\frac t\lambda \{(i+1)(aA+P_2/2+P_3/(12a)-P_5/(720a^3)) +a^2-P_2/12\}.\label{eq:low-tail-rationals} \end{align} At $a=c/\log\lambda$, \begin{align} V_i/w_i&\le V_U(a) :=-(1+cR)T_L(a)+\{1+(\lambda+1)R\}P_U(a) -2R(i+1)\lambda r^2-2Rt,\nonumber\\ V_i/w_i&\ge-(1+cR)T_U(a)+\{1+(\lambda+1)R\}P_L(a) -2R(i+1)\lambda r^2-2Rt.\label{eq:low-V-bound} \end{align} For $q\le10$, $c\le1/6$, and every integer $i\ge1$, these imply the rational bounds \begin{equation}\label{eq:low-V-rational-bound} V_L\le V_i/w_i\le V_U, \end{equation} where \[\begin{aligned} V_U&=V_U(a_U),\\ V_L&=-(1+cR)T_U(a_U)+\{1+(\lambda+1)R\}P_L(a_L)\\ &\quad-2R(i+1)\lambda r^2-2Rt. \end{aligned}\] The tail $\sum_{s>i}w_s$ lies between $w_iT_L(a_L)$ and $w_iT_U(a_U)$ on this domain. \end{lemma} \begin{proof} Geometric expansion of $\beta_s^2$ and summation over every $s>i$ give \[ \sum_{s>i}w_s=\frac{h^2}{\lambda\delta} \sum_{N\ge1}d^{N-1}\lambda^{-N(i+1)} \frac{N\delta}{1-e^{-N\delta}}. \] Hold $h,d$ fixed while differentiating this deterministic series with respect to $\lambda$. A direct derivative of its $N$th summand gives the identity \begin{align} V_i={}&-(1+cR)\sum_{s>i}w_s-2R(i+1)w_{i+1}-\frac{2Rq\beta_{i+1}}\lambda \nonumber\\ &+\{1+(\lambda+1)R\}\frac{ch^2}{\lambda^2} \Biggl\{\frac{i+1}{\delta}\sum_{N\ge1}N d^{N-1}\lambda^{-N(i+1)} \frac{N\delta}{1-e^{-N\delta}} \nonumber\\[-2pt] &\hspace{39mm}+\frac1{\delta^2} \sum_{N\ge1}d^{N-1}\lambda^{-N(i+1)} \frac{(N\delta)^2}{2(\cosh(N\delta)-1)}\Biggr\}. \label{eq:low-full-tail-identity} \end{align} The terms $-2R(i+1)w_{i+1}$ and $-2Rq\beta_{i+1}/\lambda$ come from shifting the summation indices in the terms involving $W_i^p$. For a direct check of the differentiation, write $y=\lambda^{-N}$ and $T_N=h^2N d^{N-1}y^{i+1}/[\lambda(1-y)]$. Then \[ T_N'=-T_N/\lambda- \frac{h^2N^2d^{N-1}y^{i+1}}{\lambda^2} \left\{\frac{i+1}{1-y}+\frac{y}{(1-y)^2}\right\}. \] Substitution proves~\eqref{eq:low-full-tail-identity} term by term. On a compact neighborhood of a fixed $\lambda>1$, the summands and their derivatives are bounded by a geometric sequence times a fixed polynomial in $N$. This justifies the derivative and both summations. For $p=\lambda^{i+1}$, the geometric moments are $M_j=\sum_{N\ge1}N^j d^{N-1}p^{-N}$, where $M_0=(p-d)^{-1}$ and $M_{j+1}=-p\partial_p M_j$. Applying this recurrence up to $j=5$ gives \[ \frac{hM_0}{w_i}=\frac{\lambda r}{q(b+c)},\qquad \frac{h^{j+1}M_j}{w_i} =\frac{q^{j-1}\lambda rP_j}{b+c}\quad(1\le j\le5). \] Use~\eqref{eq:low-global-F}--\eqref{eq:low-global-C} in every $N$ summand of~\eqref{eq:low-full-tail-identity}. The sign of the tail coefficient is negative and those of the two derivative kernels are positive. The resulting inequalities are precisely \eqref{eq:low-tail-rationals}--\eqref{eq:low-V-bound}. For $q\le10$ we have $c\le1/6$ and $b\le d/(q+i)$. Hence $A\le2/3$, $0\le B\le A$, $P_2\le8/9$, $P_3/A\le8/3$, and $P_4\le128/27$. Differentiation of the explicit rational functions gives \[ 00,\quad P_L'(a)>0\qquad(a\ge1). \] For instance, $T_L'/t=1-P_2/(12a^2)+P_4/(240a^4)$, and $\lambda P_U'/t=(i+1)(A-P_3/(12a^2))+2a-P_4/(120a^3)$; the displayed polynomial bounds prove these estimates. The derivative of $V_U(a)$ is positive since \[ \frac95\{1+(\lambda+1)R\}-\frac{11}{10}\lambda(1+cR)>0 \qquad(1<\lambda\le7/6, R\ge0). \] Now~\eqref{eq:low-log-enclosure} proves \eqref{eq:low-V-rational-bound} and the tail enclosure. To prove positivity, put $z=d\lambda^{-i}$ and $\Pi(N)=N(1-z)^2z^{N-1}$ for $N\ge1$. This is a probability distribution. The same geometric summation, including the two endpoints, gives \begin{equation}\label{eq:low-positive-mixture} \frac{V_i}{w_i}=\frac1\lambda\E_\Pi [i(A_N+RP_N)+B_N+RC_N], \end{equation} where, writing $Q_N=1+\lambda+\cdots+\lambda^{N-1}$, \[\begin{gathered} A_N=\frac{N}{Q_N},\quad B_N=\frac{\lambda Q_N'}{Q_N^2},\quad P_N=(\lambda+1)A_N-2\lambda^{1-N},\\ C_N=(\lambda+1)B_N- \frac{2\lambda(1-A_N)}{cN\lambda^N}. \end{gathered}\] Here $A_N>0$, $B_N\ge0$, and $P_N\ge cN\lambda^{-N}>0$: the latter is the convex-chord inequality $N(\lambda^{N+1}+1)\ge2\sum_{j=1}^N\lambda^j$. Also $Q_N-N\le cQ_N'$ term by term, and hence \[ C_N\ge B_N\left\{(\lambda+1)-\frac{2Q_N}{N\lambda^N}\right\}\ge0. \] This proves $V_i>0$ and completes the proof. \end{proof} The tail estimates also provide power-sum enclosures used below. For $A=q+2$ and every integer $j\ge2$, put \begin{equation}\label{eq:low-power-enclosures} U_j=\frac{A^{1-j}}{j-1}+\frac1{2A^j}+\frac{j}{12A^{j+1}},\qquad L_j=U_j-\frac{j(j+1)(j+2)}{720A^{j+3}}. \end{equation} Then $L_j\le\sum_{i\ge2}(q+i)^{-j}\le U_j$. Indeed, write the sum as \[ \frac1{(j-1)!}\int_0^\infty \frac{z^{j-1}e^{-Az}}{1-e^{-z}}\dd z \] and insert the global bounds~\eqref{eq:low-global-F}; Tonelli and the gamma integral evaluate every term. Thus~\eqref{eq:low-power-enclosures} includes the entire infinite tail. \subsubsection{The range \texorpdfstring{$00$. Hence $p_s\ge0$. Finally set \[ G_s=\frac{s\lambda^{s-1}(1+L_s)a_s\delta_s}{u^2},\quad G=\sum_{s\ge1}G_s,\quad G_+=\sum_{s\ge1}G_s\left\{\lambda+ \frac{(s+1)\lambda a_{s+1}}{s a_s(1+L_s)}\right\}\ge\lambda G. \] Separating the term at $s=1$ gives \begin{equation}\label{eq:low-small-raw} \frac{F_{\rm cap}}{u^2} =M\left\{P+u\left(\frac{t^2}{q_a}-2H\right) +\frac{ut}{q_a}G_+\right\}+2\alpha G. \end{equation} Here is a verification that also specifies the endpoint in this identity. Put $H_s^{\rm loc}=s-(s+1)\lambda L_{s+1}$ and $D_s^{\rm loc}=2sL_s-s/\lambda+H_s^{\rm loc}r_s$. Expanding $F_{\rm cap}$ gives \begin{align*} F_{\rm cap}={}&-2R_*uL_1 +\sum_{s\ge1}\lambda^{s-1}a_s^2 \{(1+v)W_s^0+2R_*D_s^{\rm loc}\}\\ &+2\sum_{s\ge1}s\lambda^{s-1}a_s\delta_s \left\{\alpha(1+L_s)+R_*\left(c+\lambda L_s +\frac{(s+1)\lambda L_{s+1}}s\right)\right\}. \end{align*} The elementary identities \[ W_s^0-1/\lambda=p_s-2us f_s/R_s \] and, writing $t_N=u^{-2}\sum_{s=1}^N\lambda^{s-1}a_s^2$ and $Q_N=\sum_{s=1}^N\lambda^{s-1}a_sa_{s-1}$, \begin{align*} \sum_{s=1}^N\frac{\lambda^{s-1}a_s^2}{u^2} (D_s^{\rm loc}-W_s^0) ={}&t_N-1-\frac{Q_{N+1}-u}{\lambda u^2}\\ &+\frac1{u^2}\sum_{s=1}^N\lambda^{s-1}a_s\delta_s \left\{H_s^{\rm loc}+(s+1)\lambda\frac{a_{s+1}}{a_s}\right\}\\ &+\frac{(N+1)\lambda^Na_{N+1}^2}{u^2} \end{align*} give~\eqref{eq:low-small-raw} after substitution of $R_*=ut/[2q_a-(\lambda+1)ut]$. The last endpoint tends to zero geometrically. The middle sum involving $\delta_s$ combines with the other terms proportional to $R_*$ to give the term containing $G_+$. We next bound the loss $t_\pi-t$ by $G$. With $g_i=i\lambda^{i-2}\beta_i^2$, $u^2G=\sum_i g_i\rho_i\Delta_i$. For the reference moments $\beta_i$, \[ \sum_{n>i}\lambda^{n-1}\beta_n\beta_{n-1}=q\beta_i, \qquad \sum_{n>i}w_n\alpha_p =\frac{(\lambda+1)(2-\kappa)-2\lambda u} {2-(\lambda+1)u\kappa} \ge\frac{2/3-(12/5)u}{2-2u(1-u)}\ge\frac{992}{3615} \quad(\lambda\le6/5, u\le1/16). \] For the middle inequality the numerator is positive, $\kappa\le5/3$, and $\kappa\ge1-u$; the last rational function decreases on the stated interval, since its derivative numerator is a positive constant times $18u^2-10u-13<0$. Put $B=(1-u)t_\pi^2-2H_\pi$. Equation~\eqref{eq:low-small-relative-loss} gives $t^2/q_a-2H\ge B-4\lambda t_\pi G$. Retaining $G_+$ in~\eqref{eq:low-small-raw} now yields \begin{equation}\label{eq:low-small-budget} \frac{D_\lambda}{u^2}\ge M(P+uB)+C_*G,\qquad C_* =\frac{1984}{3615} -\frac{90}{79}\frac65\frac1{16} \left(\frac{64}{9}-\frac{15}{16}\right) =\frac{389899}{18277440}>0. \end{equation} The $s=2$ term in $P$ and the $s=1$ term in $G$ give \[ P\ge\frac{c^2\rho_2^2}{(\lambda+1)^3},\qquad G\ge\frac{1-\rho_2}\lambda. \] Thus~\eqref{eq:low-small-budget} also holds in the form \begin{equation}\label{eq:low-small-root-budget} \frac{D_\lambda}{u^2}\ge M\{P+uB+K_0(1-\rho_2)\},\qquad K_0=\frac{389899}{24986880}. \end{equation} For clarity, the finite rational input to the remaining scalar estimate is stated separately. For $\ell>1$ let $c_\ell=\ell-1$, $f_s(\ell)=\ell^{s-1}c_\ell^2/(\ell^s-1)^2$, $B_s(\ell)=c_\ell/(\ell^s-1)$, and put \begin{align*} \tau(\ell)&=\frac{c_\ell^2} {\ell(c_\ell-c_\ell^2/2+c_\ell^3/3)(\ell^9-1)} +\frac{f_9(\ell)}2,\\ E(\ell)&=\frac{c_\ell^4\ell^{-16}}{6\ell^2} \left\{\frac{8}{2c_\ell/(\ell+1)} \frac{3-\ell^{-8}}{(1-\ell^{-8})^3} +\frac1{(2c_\ell/(\ell+1))^2} \left(\frac{2-\ell^{-8}}{(1-\ell^{-8})^2}-\frac12\right)\right\}. \end{align*} \begin{lemma}[Six rational bounds used for $01559/1000,\\ \sum_{s=1}^8 f_s(23/20)+\tau(23/20)&>63/40,\\ \sum_{s=1}^8 f_s(11/10)+\tau(11/10)&>19/12,\\ \sum_{s=1}^8 f_s(6/5)(1-1/s)+\tfrac89\tau(6/5)&>9/25,\\ \sum_{s=1}^8s^{-3}\{1-B_s(6/5)\}+1/128&<2/15,\\ \sum_{s=1}^8s f_s(23/20)^2+E(23/20)&<239/200. \end{align*} All denominators in these six finite expressions are positive. \end{lemma} We now estimate the tails starting at $s=9$. Both $f_s$ and $B_s$ decrease with $\lambda$; the assertion for $f_s$ follows from the increasing function $x\coth x$, and that for $B_s$ from its finite geometric denominator. As a function of real $x>0$, $f(x)$ is decreasing and convex, with \[\begin{aligned} f'&=-\log\lambda\, f\coth(x\log\lambda/2),\\ f''&=(\log\lambda)^2f \{\coth^2(x\log\lambda/2)+\tfrac12\operatorname{csch}^2(x\log\lambda/2)\}. \end{aligned}\] Convex trapezoid comparison and $\log\lambda\le c-c^2/2+c^3/3$ imply $\sum_{s\ge9}f_s\ge\tau(\lambda)$. The function $xf(x)^2$ decreases; its logarithmic derivative is less than $-3/x$. Expand its integral tail using the positive series for $(1-e^{-x\log\lambda})^{-4}$. The identities \[ \frac{\binom{j+3}{3}}{j+2}=\frac{(j+1)(j+3)}6,\qquad \frac{\binom{j+3}{3}}{(j+2)^2} =\frac{j+2-1/(j+2)}6 \] sum the two geometric parts. Keeping the $j=0$ negative subtraction and using $\log\lambda\ge2c/(\lambda+1)$ gives $\sum_{s\ge9}s f_s^2\le E(\lambda)$. Every remaining coefficient is positive, so this substitution has the stated direction. Consequently, with \[ T_0=\sum f_s,\quad H_0=\sum s f_s^2,\quad A=\sum f_s(1-B_s),\quad C_{\rm tail}=\sum s f_s^2(1-B_s), \] Lemma~\ref{lem:low-Q0-arithmetic}, $B_s\le1/s$, and $\sum_{s\ge9}s^{-3}\le1/128$ prove \[ A>9/25,\quad C_{\rm tail}<2/15,\quad H_0<241/200, \quad H_0<239/200\ \hbox{if }\lambda\ge23/20, \] together with the three lower bounds for $T_0$ on the respective intervals of $\lambda$ specified below. The general bound for $H_0$ is $\sum_{s=1}^6s^{-3}+\int_6^\infty x^{-3}\dd x=86701/72000<241/200$. The positive geometric expansion and the convex chord for $(1-uz)^{-3}$ on $0\le z\le1$ give \[ B>(1-u)(L+18u/25)^2 -2\{\overline H+(2/15)((1-u)^{-3}-1)\}. \] Use $(L,\overline H)=(19/12,241/200)$, $(63/40,241/200)$, and $(1559/1000,239/200)$ on $(1,11/10]$, $[11/10,23/20]$, and $[23/20,6/5]$, respectively. The right side decreases in $u\le1/16$; its derivative is \[ (L+18u/25)(36/25-L-54u/25)-(4/5)(1-u)^{-4}<0. \] At $u=1/16$ the three values are $973799/51840000$, $-21361/3240000$, and $-566237/16200000$. The first is positive. On the other intervals put $A_1=10/9261$, $A_2=180/79507$, and let $d_1,d_2$ be the absolute values of the last two displayed fractions. Since $c^2/(\lambda+1)^3$ increases for $1<\lambda\le6/5$, \[\begin{gathered} P\ge A_j\rho_2^2,\qquad K_0>2A_2>2A_1,\\ A_1-d_1/16=\frac{11873177}{17781120000}>0,\qquad A_2-d_2/16=\frac{1636194841}{20608214400000}>0. \end{gathered}\] For $0<\rho<1$, $A_j\rho^2+K_0(1-\rho)\ge A_j$. Equation~\eqref{eq:low-small-root-budget} proves $D_\lambda>0$. This covers $q\le1/15$, since $2/\sqrt3<6/5$. \subsubsection{Rational bounds for \texorpdfstring{$1/150$ set $\lambda=1+c$, $d=1-cq$, $u=q/(q+1)$ and $\theta_g=1-1/[6(q+1)]$. At $i=1$, use $b=d/(q+1)$ in~\eqref{eq:low-tail-rationals}, and define \begin{equation}\label{eq:low-theta-enclosures} T_{\rm lo}=u^2\{1+T_L(a_L)\},\quad T_{\rm up}=u^2\{1+T_U(a_U)\},\qquad \theta_{\rm lo,up}= \frac{T_{\rm lo,up}}{q\{2q-(\lambda+1)T_{\rm lo,up}\}}. \end{equation} In the rational expressions that follow, $R=q\theta$ and $\alpha_p$ is given by~\eqref{eq:low-reference-calibration}, with $\theta$ treated as a variable. For variables $b$ and $i$, put \begin{align} r&=(\lambda+b)^{-1},\qquad r_{\rm prev}=(1-b)/\lambda,\nonumber\\ \tau&=1-(\alpha_p+\lambda R)r+(\lambda+1)R-R/r_{\rm prev}, \nonumber\\ \mathcal B(i,b,\theta)&=\frac{2i(\alpha_p+cR)}\lambda +\frac{2iRrb}\lambda+2(i+1)Rr^2b +\frac{a_L}{\lambda}(2i-1)\tau,\nonumber\\ K_H(i,b,\theta)&=\mathcal B(i,b,\theta)-2V_U,\nonumber\\ K_J(i,b,\theta)&=V_L+\frac{a_L}{\lambda}(2i-1)\alpha_pr -\frac{a_U R}{2\lambda} \{(2i-1)(1/r_{\rm prev}-\lambda r)+2\lambda r\}, \label{eq:low-age-rational-tests}\\ i_{\rm lo}(b)&=\frac{2(d-qb)}{2b+c(d-qb)}.\nonumber \end{align} Here $V_U,V_L$ are exactly~\eqref{eq:low-V-rational-bound}, evaluated at the free $i,b,\theta$. Both $K_H$ and $K_J$ are affine in $i$ and in $\theta$ separately. We also need a lower bound for the terms at $i=1$ and the term $C_{\rm cal}e$ in~\eqref{eq:low-master}. In the next display all tail functions are at $i=1$, $b=d/(q+1)$, and $r=(q+1)/(q+2+c)$. Put \begin{align*} V_0&=-T_U(a_U)+P_L(a_L),\\ V_R&=-cT_U(a_U)+(\lambda+1)P_L(a_L)-4\lambda r^2-2r/(b+c),\\ W_0&=c/\lambda-1+2b/\lambda,\qquad W_p=cW_0+4b/\lambda -2r\{b/\lambda+2d/(q+2+c)\},\\ \alpha_0&=\{\lambda^2q+d(\lambda+1)\}/(q+1),\qquad \alpha_1=-d(\lambda+1)/(q+1),\\ B_{10}&=2\alpha_0/\lambda,\qquad B_{11}=2(\alpha_1+cq)/\lambda+2qrb/\lambda+4qr^2b,\\ E_v&=d/(q+1)^2,\qquad \varepsilon=u-\lambda\beta_2,\qquad \phi=\frac{d-c^2(q+2+c)/6}{(q+2+c)(2q+3+c)},\\ C_0&=\frac{E_v(2-E_v)}{2q} \{2u\phi+(2-h)\varepsilon/\lambda\}. \end{align*} Define, for $0\le e<1$, \begin{align} \mathcal A(e,\theta)={}&(1-e)^2(V_0+q\theta V_R) +W_0+q\theta W_p-2\theta d/\lambda+e(B_{10}+\theta B_{11}) \nonumber\\ &+e^2\left\{\frac{2\{\alpha_0+\theta(\alpha_1+\lambda q)\}} {\lambda(1-e)} -\frac32a_Uq\theta r\right\} +C_0\{\theta(2-h)-1\}e.\label{eq:low-root-rational} \end{align} Use $U_j,L_j$ from~\eqref{eq:low-power-enclosures} to define \begin{align} K_4={}&\frac{q^2(q+1)^2}\lambda \{U_3-qL_4+c(U_2-2qL_3+q^2U_4)\},\nonumber\\ K_6={}&\frac{q^4(q+1)^2}\lambda \{U_5-qL_6+c(U_4-2qL_5+q^2U_6)\},\nonumber\\ K_7={}&\frac{q^5(q+1)^2}\lambda \{U_6-qL_7+c(U_5-2qL_6+q^2U_7)\},\nonumber\\ K_{\rm sq}={}&\frac{q^4(q+1)^2}\lambda \{U_4-2qL_5+q^2U_6+c(U_3-3qL_4+3q^2U_5-q^3L_6)\}. \label{eq:low-rational-costs} \end{align} For $s=dz/(q+1)$, $e=s^2$, $0\le z\le1$, define \begin{align} \mathcal G_1(\theta)&=\mathcal A(e,\theta) -\tfrac32e(1-e)(1-s)K_4,\nonumber\\ \mathcal G_2(\theta)&=\mathcal A(e,\theta) -4e(1+s)\left\{\left(1-\frac{qs}{1-s}\right)K_6 +\frac{qs}{1-s}K_7\right\} -8e^2K_{\rm sq}.\label{eq:low-finite-budgets} \end{align} The factor $1-s$ in the denominator of $\mathcal G_2$ is strictly positive throughout the specified domain. \begin{lemma}[Rational sign tests for $00. \] For $b=dv_0/(q+2)$ and $\theta\in\{0,\theta_g\}$ one also has \[ \partial_iK_H>0,\qquad K_H(i_{\rm lo}(b),b,\theta)+\tfrac32(qb/d)^2>0. \] The latter rational expressions are used when $c>0,v_0>0$; their cleared polynomials include the closed boundary. \item On $q=1/x$, $c=hx$, $1/10\le x\le1/3$, $0\le h\le1/2$, $b=dv_0/(q+2)$, $0\le v_0\le1$, at each $\theta\in\{0,\theta_g\}$, \[ \partial_iK_H\ge0,\quad K_H(i_{\rm lo},b,\theta)+4(qb/d)^4\ge0, \qquad \partial_iK_J\ge0,\quad K_J(i_{\rm lo},b,\theta)+8(qb/d)^4\ge0. \] \item For $1/29\le q\le3$, $0\le c\le1/6$, $0\le z\le1$, both $\mathcal G_1(\theta_{\rm lo})$ and $\mathcal G_1(\theta_{\rm up})$ are strictly positive. For $q=1/x$, $c=hx$, $1/10\le x\le1/3$, $0\le h\le1/2$, $0\le z\le1$, both $\mathcal G_2(\theta_{\rm lo})$ and $\mathcal G_2(\theta_{\rm up})$ are strictly positive. \end{enumerate} \end{lemma} All the tests are on rational functions explicitly defined above. Their denominators are products of positive numerical constants and $\lambda$, $q$, $q+1$, $q+2$, $q+2+c$, $2q+3+c$, $a_L$, $a_U$, $1-b$, $\lambda+b$, $b+c$, $2b+c(d-qb)$, $1-e$, and $1-s$. The speed endpoint denominators are positive by part~(i). After substituting $b=dv_0/(q+2)$, the only additional vanishing boundary factors are $dv_0+c(q+2)$ and $(2-cq)v_0+c(q+2)$. Both are strictly positive when $c>0$ and $v_0>0$, as required for clearing denominators. For specificity, the finite coefficient operation used to check this lemma is as follows. Map each bounded variable to $y\in[0,1]$ by the affine maps in the lemma. If $P(y_1,\ldots,y_d)=\sum_\nu p_\nu y^\nu$ has coordinate degrees $D_1,\ldots,D_d$, its tensor Bernstein coefficients are \begin{equation}\label{eq:low-Bernstein-transform} b_\kappa=\sum_{\nu\le\kappa}p_\nu \prod_{j=1}^d\frac{\binom{\kappa_j}{\nu_j}} {\binom{D_j}{\nu_j}}. \end{equation} The unbounded variable $j=i-2$ in part~(ii) stays in the ordinary power basis. Positive coefficients in those variables imply the claimed sign because their basis functions are nonnegative and sum to one in each bounded variable. For the exceptional one-variable groups, the arithmetic supplement gives the polynomial and its Euclidean Sturm sequence: positive endpoint values and equal numbers of endpoint sign variations imply positivity throughout $[0,1]$. There are fifteen degree-seven groups in part~(iv) for $\mathcal G_1$, and twenty-five degree-ten groups in part~(iii) for $K_H$. Vanishing factors at zero or one are removed explicitly before applying Sturm's theorem. The remaining part~(iv) tests have $10150$ and $10850$ positive tensor coefficients. We now prove that this finite lemma suffices. Part~(i) and the tail bounds for $\sum_{s>1}w_s$ imply \[ 0<\theta_{\rm lo}\le\theta\le\theta_{\rm up}<\theta_g<1. \] For every integer $i\ge2$, Bernoulli's inequality gives $b=cd/(\lambda^i-d)\le d/(q+i)$. Also $\lambda^i=d(1+c/b)$ and $\log\lambda\le c$, $\log z\ge2(z-1)/(z+1)$ for $z\ge1$ give $i\ge i_{\rm lo}(b)$. In the definition of $B_i$ its $k$ coefficient is nonnegative, because $\tau_i>0$. Hence $S_i/w_i\ge K_H(i,b,\theta)$. For the square coefficient, expand~\eqref{eq:low-J} as \[ J_i/w_i=V_i/w_i+k(2i-1)\alpha_pr -\frac{kR}{2}\{(2i-1)(1/r_{\rm prev}-\lambda r)+2\lambda r\}. \] The bracket is positive. Since $a_L/\lambda\le k\le a_U/\lambda$, $J_i/w_i\ge K_J(i,b,\theta)$. Affinity in $\theta$ and the nonnegative derivatives with respect to $i$ in the lemma give \begin{equation}\label{eq:low-all-age-prices} \begin{cases} S_i\ge-\tfrac32w_i\beta_i^2,\quad J_i>0,&q\le3,\\ S_i\ge-4w_i\beta_i^4,\quad J_i\ge-8w_i\beta_i^4,&3\le q\le10, \end{cases}\qquad i\ge2. \end{equation} For $q\le3$, positivity of $J_i$ follows from the lower bound in \eqref{eq:low-positive-mixture}, \[ V_i/w_i>i\lambda r_i^2 \{1+R(c/\lambda+2L_{i+1})\}. \] Indeed $A_N\ge\lambda^{1-N}$ and $P_N\ge cN\lambda^{-N}$, and summing these two geometric expectations gives the display. Use $k\ge1/\lambda$ on the positive $\alpha_p$ term and $k\le1$ on its negative bracket. The result is greater than $i\lambda r^2+(2i-1)\lambda ur+qK_R$, since $\alpha_p>\lambda^2u$, $R0). \] In the integral representation of $\Delta_i$ take $\xi$ to be the integration variable and set $z=(1-\xi)^2$. Sum over $i\ge1$ and apply Jensen's inequality under $\sigma$ to each term of the mixture. Subtracting the $i=1$ term gives \begin{equation}\label{eq:low-marker-damping} \sum_{i\ge2}m_i\Delta_i \le e\sum_{i\ge2}i m_i(1-\sqrt e)^{i-1}. \end{equation} The sums are finite or positively dominated geometric sums, so Tonelli justifies both operations. Since $\rho_i\le1-e$ for $i\ge2$, this bounds the negative part of the linear terms with $i\ge2$ in \eqref{eq:low-master}. For $q\le3$, discard $(1-s)^{i-2}\le1$ in \eqref{eq:low-marker-damping}. Bernoulli and monotonicity of $y/(y-d)^{p+2}$ give \[ w_i\beta_i^p\le \frac{q^{p+2}(1+ci)}{\lambda(q+i)^{p+2}}. \] Expand $i=(q+i)-q$ and use~\eqref{eq:low-power-enclosures}, with upper bounds on positive coefficients and lower bounds on negative coefficients. For $p=2$ this gives $u^{-2}\sum_{i\ge2}i w_i\beta_i^2\le K_4$. Thus the sum of the negative parts of $\Delta_i\rho_i S_i/u^2$ for $i\ge2$ is at most $\tfrac32e(1-e)(1-s)K_4$. For $3\le q\le10$, put $t_*=qs/(1-s)\le qd/(q+h)\le1$, using~\eqref{eq:low-e-caps}. Bernoulli and the convex chord \[ (1-t)+\frac{t}{1+y}-\frac1{1+ty} =\frac{t(1-t)y^2}{(1+y)(1+ty)}\ge0 \] give \[ (1-s)^{i-1}\le\frac1{1-s} \left\{(1-t_*)+\frac{t_*q}{q+i}\right\}. \] Use the preceding Bernoulli weight bound with $p=4$. Expanding $i$ and $i^2$ in powers of $q+i$, and applying \eqref{eq:low-power-enclosures} term by term, gives the bounds \[ \frac1{u^2}\sum_{i\ge2}i w_i\beta_i^4\le K_6, \quad\frac1{u^2}\sum_{i\ge2}\frac{qi w_i\beta_i^4}{q+i}\le K_7, \quad\frac1{u^2}\sum_{i\ge2}i^2w_i\beta_i^4\le K_{\rm sq}. \] Equation~\eqref{eq:low-marker-damping}, $\Delta_i\le i e$, and \eqref{eq:low-all-age-prices} now bound the sums of the negative parts of $\Delta_i\rho_i S_i/u^2$ and $\Delta_i^2J_i/u^2$, for $i\ge2$, by $4e(1+s)\{(1-t_*)K_6+t_*K_7\}$ and $8e^2K_{\rm sq}$, respectively. It remains to bound the terms at $i=1$ and the term $C_{\rm cal}e$. Combine the terms $F_{\rm pipe}+eS_1+e^2V_1$ before estimating: \[ F_{\rm pipe}+eS_1+e^2V_1 =(1-e)^2V_1+u^2W_1(R)-2RuL_1+eB_1. \] The lower bound for the tail $V_1$ gives $V_1/u^2\ge V_0+q\theta V_R$, and $W_2^Y/(2u^2)=3k\lambda Rr_1/2\le3a_U Rr_1/2$. To estimate $C_{\rm cal}$, replace $e$ by $E_v$ only in a componentwise lower bound for the positive coefficient $C_{\rm head}(2-e)$. At $a_2=\beta_2(1-E_v)$, \[ A_0=E_v,\qquad 2-(3\lambda+1)u+\lambda(\lambda+1)a_2=(2-h)E_v. \] The remaining factors are bounded below by $u-\lambda\beta_2$ and \[ -k\log(\beta_2/u)-1+\beta_2/u\ge\phi>0. \] For the last inequality use $-\log z\ge2(1-z)/(1+z)$ and $k\ge a_L/\lambda$, then insert $\beta_2/u=(q+1)/(q+2+c)$. This yields exactly the displayed $\phi$. Its numerator is positive in both ranges of $q$: for $q\le3$ this follows from $c\le1/6$, while for $3\le q\le10$, $c=h/q$ gives $c^2(q+2+c)\le31/216<6d$. Finally, the identity \[ \frac{\kappa-1}{Z_p}=\frac{\theta(2-h)-1}{2q}>0 \] gives $C_{\rm cal}/u^2\ge C_0\{\theta(2-h)-1\}$. Thus~\eqref{eq:low-master} proves \[ F_{\rm floor}(v)/u^2>\mathcal G_1(\theta)\quad(q\le3), \qquad F_{\rm floor}(v)/u^2>\mathcal G_2(\theta)\quad(3\le q\le10). \] Both $\mathcal G_1$ and $\mathcal G_2$ are affine in $\theta$, so part~(iv) of Lemma~\ref{lem:low-Q12-arithmetic} gives positivity throughout $[\theta_{\rm lo},\theta_{\rm up}]$ in the respective ranges of $q$. Since $1/15>1/29$, this covers $1/15= 10}} For $q\ge10$ we use the simpler bound \begin{equation}\label{eq:low-large-master} F_{\rm floor}(v)>F_{\rm pipe}-eH_{\rm head}-e^2J_{\rm minus}, \quad H_{\rm head}=\sum_{i\ge1}i[-S_i]_+,\quad J_{\rm minus}=\sum_{i\ge2}i^2[-J_i]_+. \end{equation} This follows from~\eqref{eq:low-master}, $\rho_i\le1$, $\Delta_i\le ie$, and positivity of the coefficient of $e^2$ in braces and of $C_{\rm cal}e$. The next estimates will in particular show that the two negative-part sums have finite support. Put $x=1/q$, $c=hx$, $\lambda=1+hx$ and $g=q-R$. We have $0\frac{2q}{\lambda+1}-\frac1{2\lambda}. \] Here $k>2/(\lambda+1)$ follows by strict trapezoid comparison for $\int_1^\lambda t^{-1}\dd t$. Subtract the value $2q(q-1)/[(\lambda+1)q-\lambda]$ corresponding to $R=q-1$. The difference is \[ \frac{\lambda(\lambda+1)-qc^2} {2\lambda(\lambda+1)((\lambda+1)q-\lambda)}>0, \] since $qc^2=h^2/q\le1/40$. The quotient is increasing in $T_p$, so $R>q-1$; the upper bound $R0$ therefore implies $F_{\rm pipe}>M$: the inequality first proves positivity of the lower numerator, so replacing the denominator preserves the lower bound. For $a_L\le a\le a_U$, $s\ge19/20$, $\lambda+1\le41/20$, $00$. Consequently $P_M'(a)>37/10$. It suffices to check $P_M(a_L)>0$. To bound $S_i=B_i-2V_i$, set $t=i/q$, let $z=d\lambda^{-i}$, and put $w=hz/(1-z)$. Bernoulli gives $00$. For the domain \[ 03-12/q$ for $q,h>0$ with $h\le1/2$. \end{lemma} These checks have the following finite specifications. Put $A=12a_L$. The polynomial $1800A^3P_{f(h)-2x/5}(a_L)/x^3$ has constant coefficient $380160$; the sum of the absolute values of its negative coefficients, weighted by $10^{-j}2^{-k}$ for $x^jh^k$, is $65345524702337817/204800000000<320000$. For~\eqref{eq:low-Q3-price-check}, multiply $[\mathcal B-2U-\operatorname{shape}]/x+28(1+ht)/5$ by the positive denominator \[ 3(h+w)(1+x)\lambda(\lambda+1)(1-xw)(\lambda+xw)^2. \] Substitute $w=(1-h)v_0/(1+t)$ and multiply by $(1+t)^7$. The result is a polynomial of degrees $(6,8,8,7)$ in $(x,h,t,v_0)$, affine in $g$. At $g=0,1$, transform only $x,h,v_0$ by~\eqref{eq:low-Bernstein-transform} on $[0,1/10]\times[0,1/2]\times[0,1]$, keeping ordinary powers of $t\ge0$. Each endpoint has $4319$ nonzero positive coefficients, with minimum $69/280$; the remaining coefficients are zero. For~\eqref{eq:low-AJ-rational}, the numerator of $A_J^{\rm lo}-(3-12/q)$ over \[ q(q+h)^3(2q+h)(q+1)(q+2)^2(q+h+1)(q^2+3q+2h)^2 \] has $60$ nonnegative nonzero coefficients in $q,h$, including $66q^{11}$. These specifications prove the three statements by finite rational arithmetic, with every denominator positive. The first assertion gives \begin{equation}\label{eq:low-Q3-source} F_{\rm pipe}>f(h)-2/(5q). \end{equation} To apply~\eqref{eq:low-Q3-price-check} to $S_i/w_i$, we first show that for $t=i/q$, $z=d\lambda^{-i}$ and $w=hz/(1-z)$, \begin{equation}\label{eq:low-Q3-shape} \operatorname{shape}\ge\frac{9t-7}{3(1+t)}+\frac34ht. \end{equation} For $t\le1/3$, use $w\le d/(1+t)$ and $z\le d$; the difference in~\eqref{eq:low-Q3-shape} is at least $h\{5t/4-4(t-1/3)/(1+t)+1/3\}+2h^2/3\ge0$. For $t\ge1/3$ and $y=ht<3$, the exponential series, using $n!\ge2\cdot3^{n-2}$ for $n\ge2$, gives $e^y\le1+y+y^2/[2(1-y/3)]$. Since $\lambda^i\le e^y$ and $\lambda^i\ge1+y$, \[ w\ge\frac{d(1-y/3)}{1+t-y/3+yt/6},\qquad z\le\frac d{1+y}. \] After insertion, the difference is at least \[\begin{aligned} &\frac{h}{12(t+1)(y+1)((y+6)t+6-2y)}\\ &\quad\times\Bigl\{t^3(15y^2-27y+18)+t^2(81y^2+13y+108)\\ &\hspace{19mm}+t(42y^2-48y+18)+120+152y-40y^2 +\frac{48y-16y^2}{t}\Bigr\}>0. \end{aligned}\] The first and third quadratic factors have negative discriminants, the second has positive coefficients, the constant quadratic has positive endpoint values on $[0,3]$, and the last term is nonnegative. The case $y=3$ follows by continuity. For $y\ge3$, $t\ge6$; drop the nonnegative $w$ term and use $(2h+3)d\le3$ to get $\operatorname{shape}\ge2y-1/(1+y)>3+3y/4$. This proves~\eqref{eq:low-Q3-shape} in every case. Thus, with $\eta=28/(5q)$, \begin{equation}\label{eq:low-Q3-head-envelope} \frac{S_i}{w_i}\ge\frac{9t-7}{3(1+t)}-\eta +ht(3/4-\eta). \end{equation} Since $\eta\le14/25<3/4$, $S_i<0$ is possible only if $t<(7+3\eta)/(9-3\eta)$. For $i\ge2$, the lower bound from the mixture and the expression for $J_i$ give \[ J_i/w_i>iA_J-B_J, \quad A_J=\lambda r_i^2+2k\alpha_pr_i +R\{cr_i^2+2\lambda r_i^3L_i+k(\lambda r_i-1/r_{i-1})\}, \] where $B_J=k\{C_\alpha r_i-R(1/r_{i-1}-\lambda r_i)/2\}\lambda^2u$, and $R3-12/q$ by Lemma~\ref{lem:low-Q3-arithmetic}. Also $w_i\le w_1=u^2<1$, since $y/(1-dy)^2$ increases in $y=\lambda^{-i}$. Hence a finite sum bounds $J_{\rm minus}$. For $T\ge2$, writing $n=\lfloor T\rfloor$ and $r=T-n$, Faulhaber's formulas give \begin{align*} \sum_{i=1}^n i^2(T-i) ={}&T^4/12+T^2(-r^2/2+r/2-1/12)\\ &+T(2r^3/3-r^2+r/3)-r^2(1-r)^2/4 \le(T^4+T^2)/12. \end{align*} The coefficient of $T^2$ is at most $1/24$, that of $T$ is at most $1/12$ in absolute value, and $T/12\le T^2/24$ for $T\ge2$. Taking $T=(q+1)/(3-12/q)$ gives \begin{equation}\label{eq:low-Q3-square-cost} J_{\rm minus}T$ the bound on $J_i/w_i$ gives $J_i>0$. \subsubsection{Bounds for \texorpdfstring{$q\ge21$ and for $q\ge10$ with $h\ge1/4$}{q >= 21 and for q >= 10 with h >= 1/4}} We first give a quadrature bound that includes a support endpoint lying inside a mesh interval. If $G$ is continuous, compactly supported in $[0,\infty)$, $G(0)=0$, and $G'$ has bounded variation on $(0,\infty)$, then \begin{equation}\label{eq:low-BV-quadrature} \left|\Delta\sum_{i\ge1}G(i\Delta)-\int_0^\infty G(t)\dd t\right| \le\frac{\Delta^2}{8}\operatorname{TV}(G'). \end{equation} On a mesh interval $[a,b]$ use $K(s)=(s-a)(b-s)/2$ and integrate twice by parts: \[ \frac{\Delta}{2}\{G(a)+G(b)\}-\int_a^bG(s)\dd s =\int_{(a,b)}K(s)\dd G'(s). \] Derivative jumps contribute their point masses; $K$ vanishes at both mesh endpoints. Since $0\le K\le\Delta^2/8$, summing proves \eqref{eq:low-BV-quadrature}. The first and last values of $G$ are zero, so the complete trapezoid sum equals the right-endpoint sum. For $t=i/q$, $i\ge1$, monotonicity of $s/(s-d)^2$ and $\lambda^i\ge1+ht$ give \begin{equation}\label{eq:low-Bernoulli-weight} w_i\le\frac{1+ht}{\lambda(1+t)^2}. \end{equation} If $q\ge21$, use $\eta_A=4/15$, $A=39/5$, $B=41/5$ in \eqref{eq:low-Q3-head-envelope}. Every index for which $S_i<0$ satisfies $t-1/2$: after multiplying by $6(1+t)^4$ the required inequality is \[ 17-52t+36t^2+12t^3+3t^4+6\eta_A(1-t^2)>0. \] For $t\le1/3$ its first three terms are at least $11/3$; for $t\ge1/3$, use $12t^3\ge4t^2$ and $17-52t+40t^2>0$, whose discriminant is $-16$. The last group is nonnegative because $t<1$. Consequently $\operatorname{TV}(G_A')<13/5+1<4$. An elementary antiderivative gives \[ I_A:=\int G_A=\frac83-\frac{B^2}{96}-\frac B3\log(16/B). \] For $0\ell(s):=2(s+s^3/3+s^5/5+s^7/7). \end{equation} Thus, with \[ I_A^{\rm up}=175/96+(9/16)\eta_A-(3/32)\eta_A^2 -(3-\eta_A)\ell(39/121), \] equation~\eqref{eq:low-BV-quadrature} gives $eH_{\rm head} \frac16-\frac2{105}-I_A^{\rm up}-\frac1{882}-J_A =\frac{28155382207511429143103}{124404112403316879470013600} >\frac1{5000}. \] Next let $q\ge10$ and $h\ge1/4$. Use $\eta_B=14/25$, $A=217/25$, $B=183/25$. When $S_i<0$ we have $t-1$ reduces to \[ 584/25-(566/25)t+20t^2+24t^3+6t^4>0; \] the quadratic part has discriminant $-847644/625<0$. Therefore $\operatorname{TV}(G_B')<217/75+2<5$. The exact integral is $I_B=(8/3)\log(400/183)-73563/40000$. For $s_B=217/583$ the remainder of the geometric series gives \[ \log(400/183)<2\{s_B+s_B^3/3+s_B^5/[5(1-s_B^2)]\}. \] Together with~\eqref{eq:low-BV-quadrature} and $q\ge10$, this proves \[ eH_{\rm head}\frac1{12}. \] The bound for $J_{\rm minus}$ with $3-12/q\ge9/5$ gives \[ e^2J_{\rm minus}f(h)-\frac1{25}-(1-2h)C_B-(1-2h)^2J_B. \] The right side has derivative $(1-4h)/6+2C_B+4(1-2h)J_B>-1/6+2C_B>0$ on $[1/4,1/2]$. At $h=1/4$ it equals \[ \frac{16616808991576441}{1057412094912360000}>\frac1{100}. \] This proves positivity for $q\ge10$ and $h\ge1/4$, using the same $e$ in the bounds for $eH_{\rm head}$ and $e^2J_{\rm minus}$. \subsubsection{\texorpdfstring{The coefficients $S_1,\ldots,S_{24}$ for $10\le q\le21$ and $h\le1/4$}{The first 24 coefficients S(i) for 10 <= q <= 21 and h <= 1/4}} Put $x=1/q\in[1/21,1/10]$. Since $25/21>217/183$, \eqref{eq:low-Q3-head-envelope} gives $S_i\ge0$ for $i\ge25$. It remains to bound $S_i$ for $1\le i\le24$. We first bound $g$ and $F_{\rm pipe}$. With $s,s_l,P_M$ as in~\eqref{eq:low-source-rationals}, let \[ t_g(\gamma)=\frac{2(1-\gamma x)}{x+(\lambda+1)(1-\gamma x)}. \] For $0t_g(1/4),\qquad t_g(1/8)>s(a_U)/\lambda,\qquad P_{69/500}(a_L)>0. \end{equation} They imply $1/869/500$ by monotonicity of $R$ in $T_p$ and the bound using $P_M$ preceding Lemma~\ref{lem:low-Q3-arithmetic}. We next define a polynomial for each of the remaining indices. To form $F_i(c,x,g)$ take $\lambda=1+c$, $d=1-c/x$ and $a_U=1+c/2-c^2/12+c^3/24$; substitute $c=hx$ at the end. For each integer $0\le n\le25$ put \begin{equation}\label{eq:low-Q3c-Dn} D_n(c,x)=1+x\sum_{j=1}^n\binom nj c^{j-1}. \end{equation} The empty sum is zero. When $c=hx$ this is $(\lambda^n-d)/h=\beta_n^{-1}$, with its continuous value at $h=0$. For $1\le i\le24$ let $n=i+1$, $p=\lambda^n$, $b_0=1/D_n$, and set \begin{align*} m_0&=b_0,\qquad m_1=pb_0^2,\qquad m_2=p(p+d)b_0^3,\\ m_3&=p(p^2+4dp+d^2)b_0^4,\\ m_4&=p(p^3+11dp^2+11d^2p+d^3)b_0^5,\\ r&=D_i/D_{i+1},\qquad r_{\rm prev}=D_{i-1}/D_i,\\ L_i&=dx/(\lambda D_i),\qquad w_i=\lambda^{i-1}/D_i^2. \end{align*} For a free $a\ge1$, use the same $R=1/x-g$ and $\alpha_p$ as in \eqref{eq:low-large-price-rationals} and define \begin{align} \operatorname{TL}_i(a)&= \{am_0/x+m_1/2+x m_2/(12a)-x^3m_4/(720a^3)\}/\lambda, \nonumber\\ \operatorname{PU}_i(a)&=\lambda^{-2} \Bigl\{n(am_1+x m_2/2+x^2m_3/(12a))+a^2m_0/x\nonumber\\ &\hspace{30mm}-xm_2/12+x^3m_4/(240a^2)\Bigr\}, \nonumber\\ \operatorname{VU}_i(a)&=-(1+cR)\operatorname{TL}_i(a) +\{1+(\lambda+1)R\}\operatorname{PU}_i(a)\nonumber\\ &\quad-2Rn\lambda^i/D_{i+1}^2-2R/(x\lambda D_{i+1}), \nonumber\\ \operatorname{BB}_i(a)&=2i(\alpha_p+cR)/\lambda+2iRrL_i +2nR\lambda r^2L_i\nonumber\\ &\quad+(a/\lambda)(2i-1) \{1-(\alpha_p+\lambda R)r+(\lambda+1)R-R/r_{\rm prev}\}, \nonumber\\ K(t,x)&=\frac{(3-9x)t-7/3+3x}{1+t},\nonumber\\ F_i(c,x,g)&=\operatorname{BB}_i(a_U) -2\operatorname{VU}_i(a_U)/w_i-K(ix,x). \label{eq:low-Q3c-head-rational} \end{align} Each $F_i$ is an explicit rational function. To clear its denominators, set \begin{equation}\label{eq:low-Q3c-head-denominator} \mathcal D_i=720\lambda^{i+1}x^2a_U^3D_{i+1}^5 (1+x)D_{i-1}(1+ix). \end{equation} For $\gamma\in\{1/8,1/4\}$ define the polynomial \begin{equation}\label{eq:low-Q3c-head-polynomial} \mathcal P_{i,\gamma}(x,h) =x^{-2}\left.\{\mathcal D_i F_i(c,x,\gamma)\}\right|_{c=hx}. \end{equation} For each $1\le i\le24$, the arithmetic lemma checks that the numerator formed using~\eqref{eq:low-Q3c-Dn} is divisible by $x^2$. \begin{lemma}[Rational inequalities for $00$. Consequently $F_i(hx,x,g)>0$ for $1\le i\le24$, $00$. Checking the identities~\eqref{eq:low-Q3c-head-rational}--\eqref{eq:low-Q3c-head-polynomial} and these finite arrays is the entire arithmetic obligation for this lemma. Positivity of $D_n$, $a_U$, $\lambda$, $x$ and $1+ix$ proves the sign of~\eqref{eq:low-Q3c-head-denominator}. The factor $x^2$ is positive for $x>0$. Affinity of $F_i$ in $g$ then covers the whole interval between the two checked endpoints. We now use the polynomial inequalities to bound $S_i/w_i$. At $a=c/\log\lambda$, Lemma~\ref{lem:low-Lambert-tails} gives $V_i\le\operatorname{VU}_i(a)$ and $B_i/w_i=\operatorname{BB}_i(a)$. We must change this one common $a$ in both terms together. For $i\le24$, $c\le1/40$, and $n=i+1\le25$, \[ p=\lambda^n1/(1+2n/q)\ge1/6$, while $b_0\le1$. The displayed moments then give $m_3/m_1\le13$ and $m_4/m_0\le150$. Differentiating only the explicit rational functions yields, for all $a\ge1$, \[ \operatorname{TL}_i'(a)<\frac{101m_0}{100x\lambda},\qquad \operatorname{PU}_i'(a)>\frac{19m_0}{10x\lambda^2}. \] Indeed, the derivative of the first has numerator $m_0/x-xm_2/(12a^2)+x^3m_4/(240a^4)$; in the second its $n$ term is positive since $x^2m_3/(12a^2m_1)\le13/1200<1$. The remaining fourth-order ratios are at most $1/16000$ and $1/8000$, respectively. As $R>q-1/4$, $cR\le1/4$ and $\lambda\le41/40$, it follows that \[ \operatorname{VU}_i'(a)> \frac{m_0}{x\lambda^2} \left\{\frac{19}{10}(2q+1/2) -\frac{101}{100}\frac{41}{40}\frac54\right\} >3q^2m_0. \] For the last inequality, subtract $3q(41/40)^2$ from the bracket; the result is $(1037/1600)q-1101/3200\ge19639/3200>0$. We proved $\tau_i<1$ and $w_i<1$. Hence \[ \frac{\dd}{\dd a} \{\operatorname{BB}_i(a)-2\operatorname{VU}_i(a)/w_i\} <47-6q^2b_0/w_i\le47-q^2<0. \] Now $a\le a_U$ and Lemma~\ref{lem:low-Q3c-arithmetic} imply \begin{equation}\label{eq:low-Q3c-actual-head} \begin{split} S_i/w_i&\ge\operatorname{BB}_i(a)-2\operatorname{VU}_i(a)/w_i\\ &\ge\operatorname{BB}_i(a_U)-2\operatorname{VU}_i(a_U)/w_i >K(ix,x),\qquad1\le i\le24. \end{split} \end{equation} This bounds $S_i/w_i$ at every index where $S_i$ can be negative. For $i\ge25$ we already have $S_i\ge0$ by~\eqref{eq:low-Q3-head-envelope}. To bound $H_{\rm head}$, we sharpen the estimate for $w_i$. Let $y=i\log\lambda$ and $\rho=(1-h/2)/a\le1$. Directly, \[ w_i=\frac{h^2}{\lambda \{h\cosh(y/2)+(2-h)\sinh(y/2)\}^2} \le\frac1{\lambda(1+\rho i/q)^2}. \] Also \[ \lambda(1-h/2)^2-d(1+c/2)^2 =\frac{h^2}{4}(\lambda-dx^2)>0. \] Since $a\le1+c/2$, this proves $\lambda\rho^2>d$ and $\lambda\rho>d$. All coefficients of $\lambda(1+\rho t)^2-d(1+t)^2$ are therefore positive, so \begin{equation}\label{eq:low-Q3c-weight} d w_i\le(1+i/q)^{-2}\qquad(i\ge1). \end{equation} Apply $e\le d/(q+1)^2$ to~\eqref{eq:low-Q3c-actual-head}. Since $S_i\ge0$ for $i\ge25$, this gives \[ eH_{\rm head}\le\frac{u^2}{q}\sum_{i\ge1}G_x(i/q),\qquad G_x(t)=\left[\frac{t\{(7-9x)-(9-27x)t\}}{3(1+t)^3}\right]_+. \] For $0\le x\le1/10$ the support ends at $T=(7-9x)/(9-27x)<1$. With $A=7-9x$, $B=9-27x$, \[ G_x'=\frac{A-2(A+B)t+Bt^2}{3(1+t)^4},\qquad G_x''=\frac{-6A-2B+(6A+8B)t-2Bt^2}{3(1+t)^5}. \] The second derivative changes sign exactly once on the support. Furthermore $G_x'>-1/2$. The required numerator is \[ 17-52t+36t^2+12t^3+3t^4+18x(-1+8t-3t^2). \] For $t\le1/4$, the quadratic part is at least $25/4$ and the last term at least $-9/5$. For $1/4\le t\le1$, the last term is nonnegative; the same split at $1/3$ used for $G_A$ proves positivity. Including the jump at $T$ gives $\operatorname{TV}(G_x')\le7/3+1=10/3$. Thus~\eqref{eq:low-BV-quadrature} gives \[ q^{-1}\sum_{i\ge1}G_x(i/q)\le I(x)+1/240,\qquad I(x)=\frac{A(A+2B)}{6(A+B)}-\frac B3\log\frac{A+B}{B}. \] For each fixed $t$, $G_x(t)$ is the positive part of an affine function of $x$, so its integral is convex on $[0,1/10]$. Its maximum is bounded by its two endpoint values. Use \eqref{eq:low-log-four} at $s=7/25$ and $61/187$ to obtain \begin{align*} I(0)&<175/96-3\ell(7/25)=\frac{56737364551}{585937500000},\\ I(1/10)&<11407/7440-(21/10)\ell(61/187) =\frac{33081533096174793233}{297463843858303887600}. \end{align*} The second rational bound exceeds the first by \[ \frac{696221180224885514072002509}{48415339169645815039062500000}>0. \] Consequently \[ eH_{\rm head}69/500$, equation~\eqref{eq:low-large-master} therefore proves \[ F_{\rm floor}(v)>L(q):=69/500-u^2C_H-u^4J(q). \] For the last continuous comparison, ordinary differentiation yields \[ \frac{(q+1)^3}{q}L'(q) =-2C_H+\frac{q(q+1)^3}{27(q-4)^4} +\frac{q^2-2q+2}{18q(q-4)^2}. \] Both nonconstant terms decrease strictly for $q\ge10$. For the first, its logarithmic derivative is $1/q+3/(q+1)-4/(q-4)<0$; the derivative of the second is $(-q^3-6q+8)/[18q^2(q-4)^3]<0$. Thus $L'$ can change sign only from positive to negative, and $L$ has no interior minimum on $[10,21]$. Its two endpoint margins are \[\begin{gathered} L(10)-1/40=\frac{195098675960795601986623}{65597470507242819053271000}>0, \\ L(21)-1/40=\frac{713805528405149594753}{359931251068547703996000}>0. \end{gathered}\] This proves $F_{\rm floor}(v)>1/40$ for $10\le q\le21$ and $h\le1/4$. \begin{proposition}[Negativity of $v'(\lambda)$ for $1<\lambda<\min(m,2/\sqrt3)$]\label{prop:entire-low-region} For every fixed offspring law satisfying the assumptions of Theorem~\ref{thm:computer}, $v'(\lambda)<0$ for $1<\lambda<\min(m,2/\sqrt3)$. \end{proposition} \begin{proof} Lemma~\ref{lem:calibration} covers $\alpha=1$. When $0<\alpha<1$, we have $q>0$ and $00$ in the last five, and directly $D_\lambda>0$ in the first. Lemma~\ref{lem:mellin-floor} gives $D_\lambda\ge F_*>F_{\rm floor}(v)$ in all remaining cases. Since $Z>0$, the derivative formula~\eqref{eq:full-sign} gives $v'(\lambda)<0$. The finite arithmetic inputs are exactly Lemmas~\ref{lem:low-Q0-arithmetic}, \ref{lem:low-Q12-arithmetic}, \ref{lem:low-Q3-arithmetic}, and \ref{lem:low-Q3c-arithmetic}; their expressions, domains, positive denominators, and uses in the proof have all been specified above. \end{proof}