\section{The mass comparison between two and four}\label{sec:native-low} Assume $2\le\alpha\le4$ and $n\ge2$, and put $\beta=(\alpha-2)/2\in[0,1]$. The source estimates give $\max m_i/\min m_i<7\alpha/2$. We shall prove that this bound is incompatible with $F_\alpha\le0$ unless the masses are equal. The analytic comparison below holds for every real anchor greater than one. \subsection{The mean and the exact self fraction} For a real anchor $N>1$ set $x=N^2$, and tilt the joint law of Theorem~\ref{thm:joint-law} by $(X_\alpha+x\chi)^\beta$. Let \[ m=\frac{\E[x\chi(X_\alpha+x\chi)^{\beta-1}]}{\E(X_\alpha+x\chi)^\beta}, \qquad \delta=\frac{B_\alpha(1)}{B_\alpha(N)}. \] Both lie strictly between zero and one. Write $r_\alpha(N)=N(\log S_\alpha)'(N)$. Differentiation gives \begin{equation}\label{eq:raw-elasticity} N(\log B_\alpha)'(N)=2+2\beta m, \qquad r_\alpha(N)=\frac{2+2\beta m}{1-\delta}. \end{equation} We estimate the two terms together, rather than requiring $m>1/2$. Put $Z=60\chi$. Formula~\eqref{eq:chi-laplace} gives, with $a=\sqrt{15t}$, \[ \Psi(t)=\E e^{-tZ}=\frac3a(\coth a-1/a),\qquad 2t\Psi'(t)=3-3\Psi(t)-5t\Psi(t)^2. \] The bound $Z\le15E_0/\pi^2$ gives a moment-generating function near zero, so comparison of Taylor coefficients is justified. For $\mu_k=\E Z^k$ it gives \[ \mu_0=1,\qquad \mu_k=\frac{5k}{2k+3}\sum_{j=0}^{k-1}\binom{k-1}{j}\mu_j\mu_{k-1-j}. \] In particular \[ (\mu_1,\ldots,\mu_6)= \left(1,\frac{20}7,\frac{90}7,\frac{6000}{77}, \frac{4146000}{7007},\frac{5400000}{1001}\right). \] For $p(z)=1-z/6+z^2/108$, direct multiplication of these six moments gives \begin{align*} A&=\E[Zp(Z)]=9/14,& B&=\E[Zp(Z)^2]=140267/309582,\\ C&=\E[Z^2p(Z)^2]=51410/63063,& 10A^2-9B&=944/17199>0,\quad 9B-5C=1/594>0. \end{align*} Let $b=5(1+\beta)/x$. Conditioning the numerator defining $m$ on $Z$, convexity of the power $\beta-1$ and concavity of the power $\beta$ give \[ m\ge\E\frac{Z}{(b+Z)^{1-\beta}(b+1)^\beta} \ge\E\frac{Z}{b+\beta+(1-\beta)Z}. \] Here $\E X_\alpha=(1+\beta)/12$ and $\E Z=1$; the second step is weighted arithmetic--geometric mean in the denominator. Cauchy--Schwarz applied to $\sqrt{Z/d(Z)}$ and $p(Z)\sqrt{Zd(Z)}$, where $d(Z)=b+\beta+(1-\beta)Z$, now gives \begin{equation}\label{eq:six-mean} m\ge\frac{A^2}{(b+\beta)B+(1-\beta)C} >\frac{9x}{x(18-8\beta)+50(1+\beta)}. \end{equation} All denominators are positive, including at $\beta=0,1$. For the self fraction we retain $W=X_\alpha+\chi$ in its numerator. Independence and the Gamma moments give \[ \E W=\frac h{60},\qquad \Var W=\frac{35\beta+48}{25200},\qquad h=5\beta+6. \] Indeed $\Var X_\alpha=(\beta+1)/720$ follows by adding the independent variances and using $\zeta(4)=\pi^4/90$; $\Var\chi=13/25200$ follows from the second moment above. These operations on the infinite sum are justified by convergence in $L^2$ of its partial sums. Put $v=\Var W/(\E W)^2\le2/h$. H\"older interpolation gives \[ \E W\le(\E W^\beta)^{1/(2-\beta)} (\E W^2)^{(1-\beta)/(2-\beta)}, \qquad \E W^\beta\ge\frac{(\E W)^\beta}{(1+v)^{1-\beta}}. \] The second inequality also holds at zero and one by direct substitution. Jensen in the denominator of $\delta$ then gives \begin{align} \delta&\ge\frac{1}{x(1+v)^{1-\beta}[1+(x-1)/h]^\beta}\notag\\ &\ge\frac{1}{x[1+(1-\beta)v+\beta(x-1)/h]} \ge\frac{5\beta+6}{x[8+\beta(x+2)]}.\label{eq:six-self} \end{align} \begin{lemma}\label{lem:low-elasticity} For $2\le\alpha\le4$ and every real $N>1$, one has $r_\alpha(N)>2+\beta$. \end{lemma} \begin{proof} By \eqref{eq:raw-elasticity}, it suffices to prove $\beta(2m-1)+(2+\beta)\delta>0$. Write $D=x(18-8\beta)+50(1+\beta)$ and $E=8+\beta(x+2)$. The bounds \eqref{eq:six-mean}--\eqref{eq:six-self} bound this expression below by $P/(xED)$, where \[ P=\beta xE(18x-D)+(2+\beta)(5\beta+6)D. \] Its full expansion in powers of $\beta$ and $1-\beta$ is \begin{align*} P/2={}&12(9x+25)(1-\beta)^3 +20(11x+80)\beta(1-\beta)^2\\ &+(7x^2-153x+2825)\beta^2(1-\beta)\\ &+(x+10)(4x^2-50x+165)\beta^3. \end{align*} Every coefficient is positive for $x>0$: the only nonimmediate ones are \[ 7x^2-153x+2825=\frac{(14x-153)^2+55691}{28},\quad 4x^2-50x+165=\frac{(4x-25)^2+35}{4}. \] At least one basis term is positive for every $\beta\in[0,1]$. Thus $P>0$. \end{proof} \subsection{A tangent at every real anchor above one} \begin{lemma}\label{lem:low-tangent} For $n>1$, $nt>1$ and $t\ne1$, \[ T_\beta(t)=\{r_\alpha(n)-(2+\beta)\}(t-1) -\log\frac{S_\alpha(nt)}{S_\alpha(n)t^2[(1+t^2)/2]^\beta}>0. \] At $t=1$ the expression is zero. \end{lemma} \begin{proof} Use the joint law tilted at the anchor $n$, and put $Y=n^2\chi/(X_\alpha+n^2\chi)$ and $m=\E_nY$. Concavity of $s^\beta$ gives \[ \frac{B_\alpha(nt)}{t^2B_\alpha(n)} =\E_n[1+Y(t^2-1)]^\beta \le(1-m+mt^2)^\beta. \] Define \[ B_m(t)=(2m-1)(t-1)-\log\frac{1-m+mt^2}{(1+t^2)/2}, \qquad H(t)=t-\frac32+\frac1{2t^2} =\frac{(t-1)^2(2t+1)}{2t^2}. \] The raw-row contribution to $T_\beta$ is therefore at least $\beta B_m(t)$. We claim \begin{equation}\label{eq:mean-debit} B_m(t)\ge-(1-2m)_+H(t),\qquad 0\le m\le1. \end{equation} For $m\ge1/2$, apply $\log a\le a-1$ to its logarithm; the resulting lower bound is $(2m-1)t(t-1)^2/(1+t^2)\ge0$. For $m\le1/2$, differentiate with respect to $m$. Direct simplification gives \[ 2H(t)-\partial_mB_m(t) =\frac{(t-1)^2(1-m)(t+1)^2}{t^2(1-m+mt^2)}\ge0. \] Integrating from $m$ to $1/2$, where $B_{1/2}=0$, proves \eqref{eq:mean-debit}. It remains to estimate the self correction quantitatively. Write $B_\alpha(N)=cN^2f(N^2)$ with $c>0$, $f>0$, $f'\ge0$ and $f''\le0$; these signs follow from the joint representation. Put $d=B_\alpha(1)>0$ and \[ Q(N)=\log(1-d/B_\alpha(N)),\qquad N>1. \] For $z=N^2$, differentiation gives \[ N^3Q'(N)=\frac{2d}{c} \frac{(f+zf')/f^2}{1-d/(czf)}. \] The numerator is positive and nonincreasing, since its $z$-derivative is $zf''/f^2-2z(f')^2/f^3\le0$. The denominator is positive and strictly increasing. Hence $N^3Q'(N)$ strictly decreases. It follows that $Q(N)+n^3Q'(n)/(2N^2)$ increases up to $n$ and decreases after $n$. Its strict maximum at $n$ yields \[ Q(n)+nQ'(n)(t-1)-Q(nt)>nQ'(n)H(t),\qquad t\ne1. \] At the anchor, \eqref{eq:raw-elasticity} gives $q:=nQ'(n)=\delta(2+2\beta m)/(1-\delta)>0$. If $m<1/2$, Lemma~\ref{lem:low-elasticity} gives \[ q-\beta(1-2m) =\frac{\beta(2m-1)+(2+\beta)\delta}{1-\delta}>0. \] If $m\ge1/2$, $q>0$ already suffices. Adding the self remainder to \eqref{eq:mean-debit} proves $T_\beta(t)>0$ because $H(t)>0$ off one. This argument includes $\beta=0$: the raw part vanishes and the self remainder is strict. \end{proof} \subsection{Application to the finite mass distribution} For $0\le\beta\le1$ put \[ j_\beta(u)=(1+2\beta)u\log u -\beta u\log((1+u^2)/2)-(1+\beta)(u-1). \] The finite-law inequality proved in Appendix~\ref{app:escort} states that \begin{equation}\label{eq:finite-entropy} P\log\E_Uu^{1/P}+\E_Uj_\beta(u)\ge0, \qquad P=3+2\beta, \end{equation} whenever $\E_Uu=1$ and the support ratio is at most $7(1+\beta)$. For the original masses, put $t_i=1/u_i$. Then $nt_i=M/m_i>1$ and $\E_w t_i=1$, since $\E_w h(u)=\E_U[u h(u)]$. Averaging $T_\beta$ cancels the term proportional to $t_i-1$. Using $\log[(1+u^{-2})/2]=\log[(1+u^2)/2]-2\log u$ in the remaining logarithm gives \[ F_\alpha =\E_wT_\beta(1/u)+P\log\E_Uu^{1/P}+\E_Uj_\beta(u). \] If the masses are not all equal, Lemma~\ref{lem:low-tangent} makes the first term strictly positive. The support condition is exactly the bound on the maximum-to-minimum mass ratio. Equation~\eqref{eq:finite-entropy} makes the remaining sum nonnegative. Hence $F_\alpha>0$, contradicting the necessary condition at the end of Section~\ref{sec:comparison}.