\section[Channel bounds with an explicit field resource]{Channel bounds: complete operators and a resource-dependent Maxwell theorem} \label{ch:section} An efficient channel is an input pattern whose selected output retains a specified fraction of its incident power. The question here is more restrictive: every coherent combination of the counted inputs must also carry positive power through the \emph{same} section, while the absolute flux outside it stays small. Counting individual basis fields is not enough, because interference changes their flux. Nor does a count of modes of an empty aperture automatically control the fields inside an arbitrary device. The route in this section makes explicit the missing physical resource. A bound on the field in a box around the section limits how many independent inputs can each produce a prescribed through-flux. The argument first replaces the sharp section by a smooth window, then controls that window's flux by the local magnetic field. Maxwell's equations provide a derivative bound on that field, allowing a finite-dimensional approximation in the box. The resulting count depends on the local field resource; a resonator can make that resource large even when its incident power is fixed. Thus the estimate is useful when the resource is separately controlled, but it is not an area-only bound for the unrestricted class. The unrestricted fixed-section assertion is disproved later in Theorem~\ref{uc:counterexample} for $\delta>\eta$. The stronger small-bypass regime of the resource bounds below remains a separate question. This section proves three usable conclusions: an exact finite-rank reduction, a local full-vector field-resource estimate, and a bridge from the prescribed absolute-bypass condition to that estimate. The local estimates reproduce the argument in the supplied report and make its assumptions explicit \cite[Section 3]{priorreport}. The outer-window construction below avoids that report's extra conserved-collar assumption for a sharp section. None of these results bounds resonant amplification by external passivity. \subsection{What the cited channel bounds actually control} \label{ch:literature} Miller's Supplement S2 explicitly treats the maximum-index substitution as a conjecture; S12 distinguishes homogeneous-background SVD estimates from a design-specific calculation of the complete Green function. These are useful diagnostics, without a uniform theorem for every structured interior \cite[Supplements S2 and S12]{miller2023}. Kuang, Miller, and Miller first identify the physically normalized map as $R_R^{1/2}T_RGT_SR_S^{-1/2}$. Their Eq.~(1) retains the product of the two endpoint operator norms when bounding its singular values by those of $G$. Their explicit domain bounds concern the background Green operator, with source/receiver size and separation data and spherical separation; their relative-channel thresholds use a sum-rule normalization. Appendix E rejects the older largest-singular-value normalization as a general monotonic bound \cite[Section II and Appendix E]{kuang2025}. In particular, neither the endpoint norms nor these geometric data are determined by the area of the internal section. A continuum use additionally needs bounded endpoint maps: invertibility of a finite resistance matrix does not prove a uniformly bounded inverse in an infinite-dimensional current space. \subsection{Exact operator reduction} \label{ch:operators} We begin with a simple observation about the complete input-to-output map. Suppose it can be approximated, uniformly on the candidate input space, by a map with only $M$ output directions. If there were more than $M$ independent efficient inputs, some nonzero combination would disappear under the approximating map. The approximation error would then have to carry its entire output, contradicting the efficiency threshold. The following lemma states this argument without assuming that the full map is compact. Throughout, norms in the incident and delivered channel spaces are physical power norms. Let $T:\mathcal H_{\rm in}\to\mathcal H_{\rm out}$ be a bounded operator, and fix $0<\tau<1$. A counted linear subspace $W$ satisfies \begin{equation} \|Ta\|^2\geq\tau\|a\|^2\qquad(a\in W). \label{ch:throughput} \end{equation} The other passage requirements are imposed separately below. \begin{lemma}[Finite-rank obstruction] \label{ch:ranklemma} If a linear map $R:W\to\mathcal H_{\rm out}$ has rank at most $M$ and $\|T|_W-R\|<\sqrt\tau$, then $\dim W\leq M$. \end{lemma} \begin{proof} If $\dim W>M$, choose $M+1$ linearly independent vectors in $W$. The restriction of $R$ to their span has a nonzero kernel vector $a$. Then $\|Ta\|=\|(T-R)a\|<\sqrt\tau\|a\|$, contradicting \eqref{ch:throughput}. This proof also covers an infinite-dimensional $W$. The strict inequality is indispensable: on $\mathbb C^{M+1}$, $T=\sqrt\tau I$ and $R=0$ attain equality in the norm estimate. \end{proof} \begin{definition}[Approximation number] \label{ch:approxdef} For a bounded operator $U$ between Hilbert spaces, define \begin{equation} a_{M+1}(U)=\inf_{\operatorname{rank}R\leq M}\|U-R\|. \label{ch:approxnumber} \end{equation} No compactness is required for this definition. Thus $a_{M+1}(T|_W)<\sqrt\tau$ is sufficient in Lemma~\ref{ch:ranklemma}. \end{definition} \begin{proposition}[Factorization with every amplification factor retained] \label{ch:factorization} Suppose $T=AGB+E$, where all maps are bounded between their stated Hilbert spaces and $\|E\|\leq e$. If $\operatorname{rank}G_M\leq M$, then \begin{equation} \operatorname{rank}(AG_MB)\leq M,\qquad \|T-AG_MB\|\leq \|A\|\,\|B\|\,\|G-G_M\|+e. \label{ch:factorerror} \end{equation} Consequently $\dim W\leq M$ if the right side is less than $\sqrt\tau$. If $G$ is compact, one may replace $\|G-G_M\|$ by its $(M+1)$st singular value using the truncated singular-value expansion. \end{proposition} \begin{proof} The image of $AG_MB$ is contained in the image under $A$ of the at-most-$M$-dimensional image of $G_M$. Submultiplicativity and the triangle inequality prove \eqref{ch:factorerror}. Apply Lemma~\ref{ch:ranklemma}. For compact $G$, its singular expansion is an orthogonal expansion and the norm of its discarded tail is exactly its next singular value. \end{proof} \begin{remark}[Algebraic examples; no Maxwell realization] \label{ch:algebraaudit} For $G_N=\operatorname{diag}(1,2^{-1},\ldots,N^{-1})$ and $A_N=\operatorname{diag}(1,2,\ldots,N)$, the full map $A_NG_N=I_N$ is passive and has $N$ unit transmission singular values. Passivity controls the product, not $\|A_N\|=N$. Likewise, the finite-rank orthogonal projections onto the first $N$ coordinates of $\ell^2$ are individually compact contractions, although their counts above any fixed threshold below one grow without bound. Finally, compressing $\operatorname{diag}(100,1,1)$ to the last two coordinates decreases every singular value, but increases the ratio of the second to the first from $1/100$ to $1$. These exact examples identify invalid operator inferences. They supply no dielectric realization and no section or bypass certificate. \end{remark} \subsection{Local Maxwell hypotheses and a legitimate weighted flux} \label{ch:local} To apply a finite-dimensional argument to passage through a section, we need a measurement that is defined for the actual fields and can be estimated by volume norms. The smooth weight $\psi$ below measures flux through a window on the section plane; it is a mathematical aid, not a change in the physical square. A cutoff in the longitudinal direction turns this weighted plane measurement into a volume integral. The term involving the transverse gradient of $\psi$ records power crossing the side of that window, so conservation does not require us to assume that this lateral flux vanishes. Let $B\subset\mathbb R^3$ be an open rectangular box with side lengths $L_1,L_2,L_3$ and volume $V_B$. The physical dielectric extends beyond $B$; no physical boundary condition is imposed on its faces. A fixed well-posed scattering problem supplies fields depending linearly on the input $a$. Write \begin{equation} e_a=\sqrt{\varepsilon_0}E_a,\qquad h_a=\sqrt{\mu_0}H_a, \qquad c_0=(\varepsilon_0\mu_0)^{-1/2},\qquad k_0=\omega/c_0. \label{ch:normalization} \end{equation} Assume on $B$, in distributions, \begin{equation} \nabla\times e_a=ik_0h_a,\qquad \nabla\times h_a=-ik_0\varepsilon_r e_a,\qquad 0<\varepsilon_-\leq\varepsilon_r(x)\leq\varepsilon_+<\infty. \label{ch:localmaxwell} \end{equation} The real scalar coefficient may be merely measurable. Define the local weighted field resource \begin{equation} \mathcal E_B(a)=\int_B(\varepsilon_r|e_a|^2+|h_a|^2)\,dx. \label{ch:resource} \end{equation} Each field has finite resource. A uniform bound relative to incident power will be an \emph{additional} assumption. For nondispersive material this integral is four times the time-averaged local electromagnetic energy. In dispersive material a physical stored-energy formula requires frequency derivatives of the constitutive law, absent from these hypotheses. Choose a real $\psi\in W^{1,\infty}_c(\mathbb R^2)$ with $0\leq\psi\leq1$, and $\ell>0$, so that $B$ contains a neighborhood of $\operatorname{supp}\psi\times[0,\ell]$, including a negative-$z$ collar. Let $\chi\in W^{1,\infty}(0,\ell)$ be real, with $\chi(0)=1$, $\chi(\ell)=0$. Set \begin{equation} g(x,y,z)= \begin{cases} (\chi(z)\nabla_\perp\psi(x,y),\chi'(z)\psi(x,y)),&00$. Define \begin{equation} B_W=\sup_{a\in W,\,\|a\|=1}\mathcal E_B(a),\qquad \Lambda=K C_H B_W/p. \label{ch:cutoff} \end{equation} Then, with $N_{D,B}(\Lambda)$ the scalar Dirichlet eigenvalue count of the comparison box, \begin{equation} N\leq 3N_{D,B}(\Lambda) \leq\frac{V_B}{2\pi^2}\left(\frac{K C_H B_W}{p}\right)^{3/2}. \label{ch:count} \end{equation} Here $B_W$ is finite for a finite-dimensional $W$, but it is not uniformly bounded by the fixed physical class. \end{theorem} \begin{proof} Let $\Pi_\Lambda$ project onto vector Dirichlet eigenfunctions with scalar eigenvalue at most $\Lambda$. Its rank is $3N_{D,B}(\Lambda)$. If $N$ exceeds that rank, there is a unit $a\in W$ with $\Pi_\Lambda(\rho h_a)=0$. Writing $\lambda_{\rm next}>\Lambda$ for the next scalar eigenvalue, Parseval and \eqref{ch:Honebound} imply \begin{equation} \|\rho h_a\|_2^2\leq \frac{\|\nabla(\rho h_a)\|_2^2}{\lambda_{\rm next}} \leq\frac{C_HB_W}{\lambda_{\rm next}}. \label{ch:tail} \end{equation} Equation~\eqref{ch:fluxdomination} contradicts $q_\psi(a,a)\geq p$. The box eigenvalues are \begin{equation} \lambda_{\mathbf n}=\pi^2\sum_{i=1}^3 n_i^2/L_i^2, \qquad\mathbf n\in\{1,2,\ldots\}^3. \label{ch:boxspectrum} \end{equation} For each point below the cutoff, the unit cell $\prod_i[n_i-1,n_i]$ lies in the positive octant of $\sum_i x_i^2/L_i^2\leq\Lambda/\pi^2$. These cells have disjoint interiors. Their total volume is at most the ellipsoid-octant volume, giving $N_{D,B}(\Lambda)\leq V_B\Lambda^{3/2}/(6\pi^2)$. Multiplication by three proves \eqref{ch:count} without a Weyl asymptotic or a remainder depending on fine material interfaces. \end{proof} \begin{theorem}[Cost averaged over an orthonormal input basis] \label{ch:average} Under the hypotheses of Theorem~\ref{ch:resourcecount}, let $a_1,\ldots,a_N$ be any orthonormal basis of $W$, and set $\mathcal E_{\rm sum}=\sum_i\mathcal E_B(a_i)$, $B_{\rm avg}=\mathcal E_{\rm sum}/N$. If $\lambda_j^{\rm vec}$ lists the Dirichlet spectrum with three component multiplicity, then \begin{equation} \mathcal E_{\rm sum}\geq\frac{p}{K C_H} \sum_{j=1}^N\lambda_j^{\rm vec} \geq\frac{3p}{5K C_H} \left(\frac{2\pi^2}{V_B}\right)^{2/3}N^{5/3},\qquad N\leq\frac{V_B}{2\pi^2} \left(\frac{5K C_HB_{\rm avg}}{3p}\right)^{3/2}. \label{ch:averagebound} \end{equation} \end{theorem} \begin{proof} Let $A:\mathbb C^N\to H^1_0(B)^3$ take coefficients to $\rho h_{\sum_i c_i a_i}$. Its $L^2$ Gram matrix $S=A^*A$ obeys $S\geq(p/K)I$, by flux coercivity and \eqref{ch:fluxdomination}. The columns $u_i$ of $AS^{-1/2}$ are $L^2$-orthonormal. Let $D_\nabla$ be their positive semidefinite gradient Gram matrix. Cyclicity of finite matrix trace gives \begin{equation} \sum_i\|\nabla(\rho h_{a_i})\|_2^2 =\operatorname{Tr}(S D_\nabla) \geq\frac pK\operatorname{Tr}D_\nabla \geq\frac pK\sum_{j=1}^N\lambda_j^{\rm vec}. \label{ch:averageproof} \end{equation} For the last inequality, expand the $u_i$ in a vector Dirichlet eigenbasis $v_j$. The weights $w_j=\sum_i|\langle u_i,v_j\rangle|^2$ are between zero and one by Bessel's inequality and sum to $N$. Among such weights, $\sum_j\lambda_j^{\rm vec}w_j$ is minimized by filling the first $N$ positions; equivalently compare all weights with the cutoff $\lambda_N^{\rm vec}$. All series are nonnegative. The left side of \eqref{ch:averageproof} is at most $C_H\mathcal E_{\rm sum}$. The box count gives $\lambda_j^{\rm vec}\geq(2\pi^2j/V_B)^{2/3}$, and $\sum_{j=1}^Nj^{2/3}\geq\int_0^Nx^{2/3}dx=3N^{5/3}/5$. This proves both inequalities. Basis independence of $\mathcal E_{\rm sum}$ follows because it is the trace of the positive resource quadratic form restricted to $W$. \end{proof} \subsection{Absolute bypass supplies the sharp-section bridge} \label{ch:bypassbridge} The preceding theorem concerns a smooth window, whereas the question is posed for the sharp square $D$. Choose a window equal to one on $D$ and tapering to zero outside it. The correction introduced by the taper can have either sign, but its magnitude is at most the allowed absolute bypass. Passage and throughput supply the positive contribution $\eta\tau$ per unit incident power, leaving a positive margin precisely when the sufficient condition $\delta<\eta\tau$ holds. This is a condition for this bridge; it does not settle the other threshold regimes. The following trace hypothesis states explicitly how the physical section measurement is related to Lemma~\ref{ch:fluxlemma}. It is automatic for smooth fields on a neighborhood of the section, and also whenever the relevant plane Poynting integrals converge in $L^1$ to their plane-$0$ values. It is not inferred from $H(\operatorname{curl})$ alone. \begin{assumption}[Physical section compatibility] \label{ch:tracecompatibility} For each $a\in W$ there is a real flux density $s_a$ on the entire section plane, integrable on $D$ and absolutely integrable on its complement, such that \begin{equation} P_D(a)=\int_Ds_a,\qquad q_\psi(a,a)=\int_{\mathbb R^2}\psi s_a,\qquad \int_{\mathbb R^2\setminus D}|s_a|\leq\delta\|a\|^2. \label{ch:traceagreement} \end{equation} Here $s_a=\tfrac12\operatorname{Re}(E_a\times\overline{H_a})\cdot\hat z$ when that classical trace exists. A separately specified regularized measurement may be used if it establishes these identities. \end{assumption} \begin{lemma}[An outer window preserves passage up to bypass] \label{ch:outerwindow} Take $\psi=1$ on the same fixed section $D$, with $0\leq\psi\leq1$ and compact support in a slightly larger transverse region. Under Assumption~\ref{ch:tracecompatibility} and the physical passage condition $P_D(a)\geq\eta\|Ta\|^2$, one has, for every $a\in W$, \begin{equation} q_\psi(a,a)\geq\eta\|Ta\|^2-\delta\|a\|^2. \label{ch:outerwindowbound} \end{equation} If also \eqref{ch:throughput} holds and $\delta<\eta\tau$, then $q_\psi(a,a)\geq p\|a\|^2$ with $p=\eta\tau-\delta>0$. \end{lemma} \begin{proof} Because $\psi=1$ on $D$, $q_\psi(a,a)=P_D(a)+\int_{\mathbb R^2\setminus D}\psi s_a$. The second integral is at least $-\int_{\mathbb R^2\setminus D}|s_a|$. Apply the passage and bypass hypotheses, then the throughput bound. \end{proof} \begin{remark}[Scope of the outer-window argument] \label{ch:outerwindowscope} The physical section $D$ has not been enlarged: its prescribed bypass condition controls the correction outside it. The window is only an analytical device. No limit of a discontinuous indicator, no unproved zero lateral flux, and no assumption about how much flux avoids an inner edge strip are used. The new restriction $\delta<\eta\tau$ is genuine for this coercivity argument. For larger tolerances this proof yields no positive lower bound, and makes no claim of impossibility or of a counterexample. A sharp flux defined on an exceptional plane without the compatibility in Assumption~\ref{ch:tracecompatibility} is likewise outside the conclusion. \end{remark} \begin{theorem}[Approximation of the complete transmission map under stated resources] \label{ch:fulloutput} Let $W$ be a closed input subspace satisfying passage and absolute bypass, Assumption~\ref{ch:tracecompatibility}, and the preceding local Maxwell hypotheses. Suppose for some finite $B_*$, \begin{equation} \mathcal E_B(a)\leq B_*\|a\|^2\qquad(a\in W). \label{ch:uniformresource} \end{equation} For any $\Lambda\geq0$, let $L_\Lambda a=\Pi_\Lambda(\rho h_a)$, let $\mathcal K_\Lambda=\ker L_\Lambda$, and let $P_\Lambda$ be the orthogonal projection of $W$ onto $\mathcal K_\Lambda^\perp$. The operator $R_\Lambda=T|_W P_\Lambda$ satisfies \begin{equation} \operatorname{rank}R_\Lambda\leq3N_{D,B}(\Lambda),\qquad \|T|_W-R_\Lambda\|^2 \leq\frac1\eta\left(\delta+ \frac{K C_H B_*}{\lambda_{\rm next}}\right), \label{ch:fulloutputerror} \end{equation} where $\lambda_{\rm next}$ is the first scalar box eigenvalue exceeding $\Lambda$. In particular, for $p=\eta\tau-\delta>0$ and $\Lambda=K C_HB_*/p$, the norm error is strictly below $\sqrt\tau$. If every vector in $W$ also satisfies \eqref{ch:throughput}, then \begin{equation} \dim W\leq\frac{V_B}{2\pi^2} \left(\frac{K C_HB_*}{\eta\tau-\delta}\right)^{3/2}. \label{ch:physicalcount} \end{equation} \end{theorem} \begin{proof} The resource bound makes $a\mapsto\rho h_a$ bounded, so $\mathcal K_\Lambda$ is closed. Since $L_\Lambda$ has rank at most $3N_{D,B}(\Lambda)$, its kernel has codimension at most that number; this proves the rank statement. For $b\in\mathcal K_\Lambda$, the spectral-tail estimate and Lemmas~\ref{ch:caccioppoli} and \ref{ch:outerwindow} give \begin{equation} \eta\|Tb\|^2\leq q_\psi(b,b)+\delta\|b\|^2 \leq\left(\frac{K C_HB_*}{\lambda_{\rm next}}+\delta\right)\|b\|^2. \label{ch:kerneloutput} \end{equation} Apply this to $b=(I-P_\Lambda)a$, whose norm is at most $\|a\|$. Because $\lambda_{\rm next}>\Lambda=K C_HB_*/p$, the error is strictly below $\sqrt\tau$ (if $B_*=0$, positivity of $p$ already precludes a nonzero counted subspace). Lemma~\ref{ch:ranklemma} and the explicit box count finish the proof. \end{proof} \begin{corollary}[Conditional uniform family bound] \label{ch:spectralcor} Let $V_\tau=\operatorname{Ran}\mathbf1_{[\tau,\infty)}(T^*T)$. If passage, absolute bypass, trace compatibility, common local geometry, and a common finite resource $B_*$ hold for every vector of $V_\tau$, with $\delta<\eta\tau$, then \eqref{ch:physicalcount} bounds $\dim V_\tau$, without an a priori compactness assumption on $T$. It is the usual count of squared singular values at least $\tau$ when $T$ is compact. Alternatively a uniform bound on the basis-average resource over every finite-dimensional subspace gives the last bound in \eqref{ch:averagebound}, with $p=\eta\tau-\delta$. \end{corollary} \begin{proof} The spectral theorem implies \eqref{ch:throughput} on $V_\tau$. Apply Theorem~\ref{ch:fulloutput}, or apply Theorem~\ref{ch:average} to every finite-dimensional subspace. An infinite-dimensional Hilbert space contains finite orthonormal sets of arbitrarily large cardinality, so a common finite upper bound forces finite dimension. \end{proof} The map $R_\Lambda$ is defined using the complete $T$, including all matching and resonant feedback. It is not obtained by deleting a component of an isolated link and then silently applying an uncontrolled inverse. Its cost is the explicitly retained norm $B_*$ of the actual full field. Thus Theorem~\ref{ch:fulloutput} discharges the output-approximation step under a resource hypothesis; it does not discharge that hypothesis. \begin{corollary}[Essential-norm consequence] \label{ch:essential} Under Theorem~\ref{ch:fulloutput}'s hypotheses, without assuming a throughput lower bound, the distance of $T|_W$ from the compact operators is at most $\sqrt{\delta/\eta}$. In particular, exact zero bypass makes $T|_W$ compact. \end{corollary} \begin{proof} The finite-rank $R_\Lambda$ are compact and $\lambda_{\rm next}\to\infty$ as $\Lambda\to\infty$. Take the limit superior in \eqref{ch:fulloutputerror}. When $\delta=0$, these finite-rank operators converge to $T|_W$ in operator norm. A norm limit of compact operators is compact: given any bounded input sequence, use a diagonal subsequence convergent under each approximant, then the uniform approximation makes its images under $T$ Cauchy. \end{proof} \subsection{An explicit area scaling with its extra resource exposed} \label{ch:explicitconstants} \begin{corollary}[Conditional square-enclosure area scaling] \label{ch:arearesource} Suppose $D\subset[-L/2,L/2]^2$ and there is a source-free comparison box \begin{equation} B=(-L/2-4r,L/2+4r)^2\times(-3r,4r),\qquad r>0, \label{ch:explicitbox} \end{equation} on which the material bounds and \eqref{ch:uniformresource} hold. There are admissible cutoffs with \begin{equation} V_B=7r(L+8r)^2,\quad G\leq\sqrt3/r,\quad d_1,d_2\leq\sqrt3/r,\quad C_H\leq k_0^2\varepsilon_++3/r^2,\quad K\leq\frac{c_0\sqrt3}{2r\sqrt{\varepsilon_-}} \sqrt{2+\frac{12}{k_0^2\varepsilon_-r^2}}. \label{ch:explicitconstantsvalue} \end{equation} In particular, if $r=(k_0\sqrt{\varepsilon_+})^{-1}$ is admissible and $\kappa=\varepsilon_+/\varepsilon_-$, then, with $p=\eta\tau-\delta>0$, \begin{equation} \dim W\leq\frac{7}{2\pi^2} (k_0\sqrt{\varepsilon_+}L+8)^2 \left(2\sqrt3\sqrt\kappa\sqrt{2+12\kappa}\, \frac{\omega B_*}{p}\right)^{3/2}. \label{ch:arearesourcebound} \end{equation} For a scaled fixed shape with $L^2\leq c_D|D|$, this is $O(1+k_0^2\varepsilon_+|D|)$ \emph{only if} $\omega B_*$, the material ratio, thresholds, shape constant, and the required collar remain controlled. \end{corollary} \begin{proof} For a one-dimensional interval use the function equal to one on that interval, decreasing linearly to zero across an exterior strip of width $r$, and zero beyond the strip. Its Lipschitz constant is $1/r$. Let $Q_s=[-L/2-s,L/2+s]^2$. Take $\psi$ as the product of the two one-dimensional cutoffs equal to one on $Q_0$ and zero outside $Q_r$, and take $\chi(z)=1-z/r$ on $(0,r)$. Then $|\nabla_\perp\psi|\leq\sqrt2/r$ and $|g|\leq\sqrt3/r$. Take $\zeta$ as the product cutoff equal to one on $Q_r\times[0,r]$, supported on $Q_{2r}\times[-r,2r]$; take $\rho$ equal to one on the latter box and supported on $Q_{3r}\times[-2r,3r]$. Each has gradient bounded by $\sqrt3/r$, and both supports are compact in \eqref{ch:explicitbox}. Equality on the indicated closed supports suffices in the preceding estimates. This proves \eqref{ch:explicitconstantsvalue}. Substitute $r=(k_0\sqrt{\varepsilon_+})^{-1}$ in \eqref{ch:physicalcount}, using $\omega=c_0k_0$; straightforward cancellation gives \eqref{ch:arearesourcebound}. Finally $(x+8)^2\leq2x^2+128$ proves the asserted order with every additional dependency retained. \end{proof} \begin{corollary}[Local extension to real symmetric anisotropy] \label{ch:anisotropy} All local estimates and their conditional channel consequences remain valid for measurable real symmetric tensors $\varepsilon_r(x)$ satisfying $\varepsilon_-I\leq\varepsilon_r\leq\varepsilon_+I$, provided $\varepsilon_r|e|^2$ is replaced by $\overline e\cdot\varepsilon_r e$ in the resource. \end{corollary} \begin{proof} Symmetry and reality give the same cross-Poynting cancellation. The magnetic estimate uses $|\varepsilon_r e|^2\leq\varepsilon_+ \overline e\cdot\varepsilon_r e$. In the Caccioppoli step use the positive quadratic form of $\varepsilon_r^{-1}$, its Cauchy--Schwarz inequality, and $\varepsilon_r^{-1}\leq\varepsilon_-^{-1}I$. The identity $k_0^{-2}\overline{\operatorname{curl}h}\cdot \varepsilon_r^{-1}\operatorname{curl}h= \overline e\cdot\varepsilon_r e$ replaces its scalar version. Every remaining step is unchanged. Constant scalar permeability is still essential to this proof of $\operatorname{div}h=0$. \end{proof} \subsection{The scope of the resource route} \label{ch:open} \begin{remark}[Open question: removing the resource in the small-bypass regime] \label{ch:missing} There is no estimate here, from $1\leq\varepsilon_r\leq\varepsilon_+$, the fixed section, passivity, passage, and absolute bypass alone, that uniformly bounds $B_*$ or replaces its role by a uniform approximation estimate for the full $T$. A possible sufficient missing theorem is: every admitted counted subspace has a finite-rank approximation with error strictly below $\sqrt\tau$ and rank bounded solely by the frozen parameters. An area scaling additionally requires a rank of order $1+k_0^2\varepsilon_+|D|$. Theorem~\ref{ch:fulloutput} supplies such an approximation \emph{with} $B_*$; it does not eliminate it. Even proving unbounded local amplification would only invalidate that proof route, not construct an arbitrarily large channel subspace. Conversely, pointwise compactness for each device would not make the bound uniform over a sequence of devices. Theorem~\ref{uc:counterexample} later disproves the unrestricted assertion in the allowed range $\delta>\eta$; it leaves the stricter regime $\delta<\eta\tau$ used here unresolved without resources. \end{remark}