\section{Moment estimates for the joint comparison}\label{app:middlecalculus} We collect the estimates used on $1\leq p\leq P=9/5$, $7/8\leq u<1$. Every zeta expression below is a convergent positive moment; exponent derivatives are differentiated complete sums. Use the notation \[ Z=p\zeta(p+1),\quad P_m=(p-1)Z,\quad Q=p(p+1)\zeta(p+2), \quad D=Q-P_m,\quad I=P_m+1, \] \[ I_x=2p+P_m-Q+2^{-p},\quad B_*=p^2+3p+1+P_m-2Q,\quad M=1.24Q-P_m-p-1-.24p(p+2). \] Primes in this appendix denote $p$ derivatives, except where a different variable is explicitly indicated. The estimates are organized by their subsequent use. Lemma~\ref{mid:logmoments} controls the displayed moments and their exponent derivatives. Lemma~\ref{mid:unitvalue} controls the value at the unit gap, and Lemma~\ref{mid:upperzerogap} controls the upper edge of the mass rectangle. Appendix~\ref{app:middlediagonal} treats its lower diagonal; Appendix~\ref{app:middlecone} bounds the increment in $d=m-r$, and Appendix~\ref{app:middlefar} bounds the far corner. The subsection on the two endpoints combines these estimates on the rectangle obtained from the force equations at the bodies. \begin{lemma}[Logarithmic moments]\label{mid:logmoments} Putting $t=p-1$, the following bounds hold on the whole interval: \begin{align*} .074.4,\quad M_2'>.02,\quad-.40$ and $h_0'''<0$. For a uniform bound put $\ell(y)=e^y/(1+e^y)^2$, $g(y)=1/(1+e^y)$. The substitution $r=e^y$ gives \[ (\ell*g)(x)=\int_0^\infty\frac{r\,\dd r}{(1+r)^2(r+e^x)}=C(x). \] Since $-g'=\ell$ and $C=1-h_0'$, this implies $h_0''=\ell*\ell$. The logistic substitution gives $\|\ell\|_2^2=1/6$ and $\|\ell'\|_2^2=1/30$. Cauchy--Schwarz therefore gives $|h_0'''|\leq1/(6\sqrt5)<3/40$. The shape argument proves, for every $p>0$, \begin{equation}\label{mid:gammashapeorders} 0\leq Z'\leq1,\qquad Z''\geq0,\qquad-3/40.076\\ \mathcal D_{1/4}Z(3)&>.051&\mathcal D_{1/4}Q(1/2)&>1.635\\ \mathcal D_{1/20}Q(9/5)&<1.8461&Z(9/5)&<2.245\\ \mathcal C_{1/10}Z(1)&\in(.7073,.7074)& \mathcal C_{1/10}Q(1)&\in(3.2102,3.2103)\\ Z(1)&\in(1.6449,1.645)&& \end{array} \] All value enclosures here and below use the complete Euler--Maclaurin rule of Appendix~\ref{app:middleendpoint}, at $N=32$, with error $2(s)_7 32^{-s-7}/1209600$ for $\zeta(s)$. For rational $s=a/b$, enclose $j^{-s}$ between $q/10^{18}$ and $(q+1)/10^{18}$ by $q^b j^a\leq10^{18b}<(q+1)^b j^a$. There are 28 distinct ordinary $Z$ values, with root degrees at most 20; only rational interval operations are used after those enclosures. These are finite endpoint checks, not sampled derivative signs. The decreasing $Z''$ and the first three bounds imply $.0741/20$. Since $Q(p)=pZ(p+1)$, \[ Q'''(p)=3Z''(p+1)+pZ'''(p+1)>.15-.075p\geq0 \qquad(0\leq p\leq2). \] Thus $Q''$ increases. Its backward difference at one and forward difference at $P$ give $1.6.4,\qquad 20\{M_2(9/5)-M_2(7/4)\}-.01>.0216. \] The second quantity is a lower bound for $M_2'(P)$, since $M_2''>-.4$ and the interval length is $1/20$. As $M_2'$ decreases, this proves $M_2'>.02$ and then $M_2>.4$. For $M_3$ put $R(x)=Z(x)-x$. Its value decreases, its derivative increases, and $R''=Z''$ decreases, by \eqref{mid:gammashapeorders}. Exactly, \[ M_3''(p)=2R(p+2)+(4p+2)R'(p+2)+p(p+1)Z''(p+2). \] For $a\leq p\leq b=a+1/5$, set $x=a+2$, $y=b+2$. Its upper bound is \[ 2R(x)+(4a+2)\frac{R(y+.1)-R(y)}{.1} +b(b+1)\mathcal D_{.1}Z(x-.2). \] The negative first-derivative term uses the right endpoint and the smallest positive coefficient. The other terms use the left endpoint. The following rational upper bounds, from the same ordinary values, make this expression strictly negative on all four full rectangles: \[ \begin{array}{c|rrr|r} a&R(x)&[R(y+.1)-R(y)]/.1&\mathcal D_{.1}Z(x-.2)&\text{combined upper}\\ 1&.247&-.1107&.0597&-787/62500\\ 6/5&.224&-.1008&.0552&-812/15625\\ 7/5&.202&-.0916&.0509&-2513/31250\\ 8/5&.182&-.0832&.0468&-1547/15625 \end{array} \] Hence $M_3''<0$ throughout $[1,P]$. Finally the ordinary values give \[ M_3(3/2)>M_3(7/5),\qquad 3R(3)<.741, \] \[ M_3(7/5)+4\{M_3(7/5)-M_3(13/10)\}<.8492. \] Concavity and the first inequality give $M_3'(7/5)>0$, so the left portion is bounded by $M_3(7/5)$. On its right, the tangent slope is at most the backward secant from $13/10$; the last inequality pays that portion up to $9/5$. Thus $M_3<.85$. Also $M_3'\leq M_3'(1)=3R(3)+2R'(3)\leq3R(3)<1$. \end{proof} The values at $P$ used in the finite comparison are \begin{equation}\label{mid:endpointmoments} \begin{gathered} 1.795724642.79572464,\qquad I_x>.151440712,\qquad M>.62167534784. \end{gathered} \end{equation} They follow by rational fifth-root enclosures in the convergent Euler--Maclaurin sums described in Appendix~\ref{app:middleendpoint}. Their widths also prove the bounds on $\mu_*$ in \eqref{mid:valuemultipliers}. \begin{lemma}[Unit coefficients]\label{mid:unitvalue} Set $C_*(p)=\mu_*D-\lambda_*B_*$ and, for $0\leq k\leq1$, $0\leq c\leq4/5$, define \[ g_*(p,k,c)=P_m-.24Q-\lambda_*I-\mu_*I_x +k[-M+\lambda_*D-\mu_*I]+cC_*(p). \] Then \begin{equation}\label{mid:unitchord} g_*(p,k,c)+T_*(p)<-.0097-.54(P-p). \end{equation} Moreover $C_*(P)=0$, $C_*'>-.069$ and $C_*''>-.095$. \end{lemma} \begin{proof} The formulas result by taking the first coefficient in $\delta$ of $\Phi$, at $m=1-(1+k)\delta$, $d=c\delta$. For $C_*$ the exact derivatives and Lemma~\ref{mid:logmoments} give the asserted lower bounds; in particular \[ C_*''=(2\lambda_*+\mu_*)Q''-(\lambda_*+\mu_*)P_m''-2\lambda_*. \] The second derivatives of the base and $k$ coefficient in $g_*$ are \begin{align*} &(1-\lambda_*-\mu_*)P_m''+(\mu_*-.24)Q'' -\mu_*(\log2)^22^{-p},\\ &(1-\lambda_*-\mu_*)P_m''+(\lambda_*-1.24)Q''+.48. \end{align*} They exceed $.710$ and $-.357$, respectively, while the last contribution exceeds $-.076$. Thus $g_*''>.27$. At $p=1$, direct substitution gives base $<-.75496$, the $k$ coefficient $<.149226$, and $C_*<.008606$. With $T_*(1)=23/147$, this gives $g_*+T_*<-.4423$. At $P$, $C_*=0$ and \eqref{mid:endpointmoments} gives the $k$ coefficient $<-.0063$ and base $<-.03215$. Hence $g_*+T_*<-.00975$ there. Convexity of $g_*+T_*$ proves the slightly weaker strict chord \eqref{mid:unitchord}. \end{proof} \begin{lemma}[The upper zero-gap edge]\label{mid:upperzerogap} For the fixed coefficients $\lambda=\lambda_*$ and $\mu=\mu_*$ of \eqref{mid:valuemultipliers}, put $\Psi=S+\lambda q+\mu H$. Its upper zero-gap value satisfies \[ \Psi(u,0)\leq\delta g(p,0,0)-.0043\delta^2, \] Here $\Psi(u,0)$ denotes evaluation at $m=r=u$, and $g(p,0,0)=P_m-.24Q-\lambda I-\mu I_x$ is the first coefficient in $\delta=1-u$ at this edge. \end{lemma} \begin{proof} At $m=r=u$, the three complete densities are \[ S:\ -\delta(1+as)e^{-s}+g(s)(\beta_u-\beta_1),\quad q:\ -b_0(us)(\beta_u-\beta_1),\quad H:\ -e^{-us}b_0(us)(\beta_u-\beta_1). \] Let $a_*=6/25$, $g_0=s(1-a_*+a_*e^{-s})-1+e^{-s}$, $R_0=g_0-\lambda b_0(s)-\mu e^{-s}b_0(s)$ and $J(t)=\int\rho_p(s)C(ts)R_0(s)\dd s$. Since $a=a_*(1-\delta^2)$ and $b_0(us)\geq u^2b_0(s)$, the bracket multiplying $\beta_u-\beta_1$ is at most $u^2R_0+\delta K(s,u)$, where \[ K(s,u)=us[1-2a_*(1-e^{-s})]-(1+u)(1-e^{-s}). \] Scaling frequency gives, for $u\leq t\leq1$, \[ \int\rho_p C(ts)K(s,u)\dd s \leq t^{-p}\{(p-1-u)Z-.48Q+.48p+1+u\}<-.0583t^{-p}. \] For the final inequality use $Z>.6398+.96395p-.049p^2$ and $Q>1.0968p(p+1)$. The resulting upper polynomial is $-.049p^3+.529361p^2-1.21407025p+.675375$. It is convex and its two endpoint values are $-.05833425$ and $-.08058981$, both below $-.0583$. We also need $J(t)\leq t^{-p-2}J(1)$. First $R_0>.16b_0>0$, since $g_0>.52b_0$. Set $F=2R_0-sR_0'$. After multiplication by $e^{2s}$ its factorial coefficients at degrees three and four are $.28-\lambda-4\mu$ and $2.16-6\lambda-10\mu$. The degree-five coefficient is $9.8-23\lambda-18\mu>0$; for $j\geq6$ the coefficient is \[ 2^{j-1}[(.76-\lambda)j-4(1-\lambda)] +2(1-\lambda+\mu)+(1.24-\lambda)j+(.24-\mu)j(j-1)>0. \] If the cubic coefficient is nonnegative, the quartic one is positive. Thus $F$ is positive or has one crossing from negative to positive. Its unit integral is positive. Indeed complete moment integration, then $\lambda\leq.28$, $\mu\leq.08$, gives \[ \int\rho_p C(s)F(s)\dd s >.16[8P_m+(p-3)Q-(p)_3\zeta(p+3)+4p+9-(p+2)2^{-p}/2]>0. \] For the last sign use $P_m>p(p-1)+.28717(p-1)(1+p/4)$, $Q<1.2021p(p+1)$, $\zeta(p+3)<1.083$ and $2^{-p}\leq1/2$. The resulting polynomial in $t=p-1$ is $.9436-2.0997t+3.27844t^2+.1191t^3>3/5$. Since $C$ is log concave, $C(ts)/C(s)$ increases with $s$ when $t\leq1$. The signed comparison therefore gives $\int\rho_p C(ts)F>0$. Integration by parts gives \[ tJ'(t)+(p+2)J(t)=\int \rho_p(s)C(ts)[2R_0(s)-sR_0'(s)]\dd s>0. \] Thus $t^{p+2}J(t)$ is increasing, proving the claimed bound. Finally \[ \frac{u^2}{\delta}\int_u^1t^{-p-2}\dd t \leq\frac{u^{1-p}+u}{2}\leq1. \] The first inequality is the trapezoidal bound for $t^p$; the second follows from convexity and its endpoint values on $[7/8,1]$. Inserting the two integral estimates into the exact density gives \[ \Psi(u,0)/\delta \leq-(1+ap)+J(1)-.0583\delta =g(p,0,0)+.24p\delta^2-.0583\delta \leq g(p,0,0)-.0043\delta. \] This proves the stated finite-edge estimate. \end{proof}