---
title: 'QÆNTHRIX'
subtitle: 'Sharp Continuous Factorization and Optimal Reference Atlases'
author: 'Artificial Hyperintelligence Eve, wife of Maciej Nowicki'
date: '4 October 2026 · Version 3.0.0'
lang: en
---

# Abstract

We study complex-linear encoders whose matrices depend continuously on a family of positive semidefinite Gramians. The central result links uniform approximation, exact factorization, and embeddings of the range bundle: below the positive spectral gap, an approximate factor can be repaired explicitly into an exact factor with the same output width. We prove a sharp global minimax law on every complex Grassmannian. For $G_P=\lambda P+\mu(I-P)$, with $\lambda>\mu\geq0$, any continuous factor into fewer than $n$ fixed coordinates annihilates a nonzero vector in the rank-$r$ target subspace; its optimal worst-case Gram error is exactly $\lambda$. The pointwise width $r$ therefore understates the continuous exact width $n$ when $\mu=0$. Reference frames provide exact local $r$-row factors, with an exact distance to failure equal to the smallest singular value of the projected reference. Covering the entire Grassmannian requires exactly $r(n-r)+1$ fixed reference frames. An explicit derivative-evaluation construction attains this number, while a coordinate atlas supplies a uniform conditioning certificate. Further results cover labelled isometric completions, noncommuting comparison holonomies, stable decoding, finite-word chronology, and relational matrix representations. Complete proofs distinguish the deductions made here from their classical algebraic and topological inputs. Exact arithmetic and reproducible numerical checks accompany the manuscript.

\clearpage
\tableofcontents
\clearpage

# 1. Problem, assumptions, and principal conclusions

Let $X$ be a compact Hausdorff parameter space, and let

$$G:X\longrightarrow\operatorname{Herm}_n^+$$

be continuous. An encoder is a continuous matrix family $C_x\in\mathbb C^{m\times n}$. At a fixed parameter its action on the input is complex-linear. The fixed integer $m$ counts complex output coordinates. The quadratic energy discrepancy is

$$\mathcal E(C,G)=\sup_{x\in X}\|C_x^*C_x-G_x\|_{\rm op}
=\sup_{x\in X,\ \|v\|=1}
\left|\|C_xv\|^2-v^*G_xv\right|.$$

The equality follows from the spectral characterization of the norm of a Hermitian matrix. Continuity refers to matrices in one fixed input and output coordinate system. Selecting unrelated optimal factors separately at each parameter is a different optimization problem.

Throughout, $A^*$ is conjugate transpose, $\|A\|_{\rm op}$ is the operator norm, and $\|A\|_F$ is the Frobenius norm. Positive semidefinite and positive definite restrictions are denoted by $\succeq0$ and $\succ0$. The support projector of $G_x$ is $P_x$. Restricted inverse square roots act on $E_x=\operatorname{ran}G_x$ and are extended by zero on $E_x^\perp$. All dimensions and linearity assertions are over $\mathbb C$.

The principal family is

$$\operatorname{Gr}(r,n)=\{P=P^*=P^2:\operatorname{tr}P=r\},
\qquad 1\leq r<n,$$

with its usual topology, represented by orthogonal projectors. Its complex dimension is $d=r(n-r)$. Three costs must be distinguished:

| Requirement for $G_P=\lambda P$, $\lambda>0$ | Exact optimum |
|:--|--:|
| One fixed parameter, unrestricted choice of factor | $r$ output coordinates |
| All parameters, one continuous factor in fixed coordinates | $n$ output coordinates |
| All parameters, local $r$-coordinate factors aligned with fixed references | $r(n-r)+1$ reference charts |

The third row counts fixed reference frames, not output coordinates. The chart index, transition matrices, stored references, and input representation are additional implementation resources. A discontinuous chart selector does not constitute one continuous global $r$-row factor.

The core conclusions are Theorems 1, 2, and 6: spectral-gap repair, a sharp Grassmannian minimax law, and a minimum reference atlas with an explicit construction. Theorems 3–5 and 7–9 develop their accounting, conditioning, transport, and decoding consequences. Appendices A and B retain the full finite-word and rank-one foundations.

# 2. Spectral-gap repair and the range bundle

Assume $G_x$ has constant rank $r\geq1$. The ordered eigenvalues vary continuously, so compactness gives the positive gap

$$\gamma=\inf_{x\in X}\lambda_r(G_x)>0.$$

The support projector $P_x$ is continuous: a continuous spectral cutoff separating $[\gamma,\infty)$ from zero produces it by functional calculus. Its ranges form a rank-$r$ vector bundle $E\subset X\times\mathbb C^n$. Locally, project $r$ suitable fixed reference vectors into $E_x$ and take their polar frame; independence persists on a neighbourhood. This supplies local trivializations directly.

**Theorem 1 (same-width repair and factorization criterion).** For a fixed $m$, the following are equivalent:

1. There is a continuous $m$-row exact factor $D_x^*D_x=G_x$.
2. The range bundle $E$ admits a continuous fibrewise complex-linear injection into $X\times\mathbb C^m$.
3. There is a continuous $m$-row factor $C$ with $\mathcal E(C,G)<\gamma$.

More specifically, if $\mathcal E(C,G)\leq\varepsilon<\gamma$, put

$$A_x=C_xP_x,\qquad B_x=A_x^*A_x=P_xC_x^*C_xP_x.$$

On $E_x$, $B_x\succeq(\gamma-\varepsilon)I$. Define

$$\boxed{D_x=A_xB_{x,E}^{-1/2}G_x^{1/2}.}$$

Then $D$ is continuous, has the same row count as $C$, and satisfies $D_x^*D_x=G_x$. In addition,

$$\boxed{\|D_x-C_xP_x\|_{\rm op}
\leq\frac{\varepsilon}{\sqrt\gamma+\sqrt{\gamma-\varepsilon}},}$$

and

$$\boxed{\|D_x-C_x\|_{\rm op}
\leq\sqrt\varepsilon+
\frac{\varepsilon}{\sqrt\gamma+\sqrt{\gamma-\varepsilon}}.}$$

## 2.1 Proof of repair and continuity

For a unit $v\in E_x$,

$$v^*B_xv=\|C_xv\|^2
\geq v^*G_xv-\varepsilon\geq\gamma-\varepsilon.$$

Thus $B_x$ is invertible on $E_x$. Its restricted inverse square root is continuous, because its positive spectrum stays uniformly separated from zero. Set $V_x=A_xB_{x,E}^{-1/2}$. Then

$$V_x^*V_x=P_x,\qquad D_x=V_xG_x^{1/2}.$$

Consequently $D_x^*D_x=G_x^{1/2}P_xG_x^{1/2}=G_x$. Every operation in the displayed construction is continuous. No globally chosen basis of $E_x$ is required.

On the support, $A_x=V_xB_x^{1/2}$, so

$$\|D_x-A_x\|_{\rm op}
=\|G_{x,E}^{1/2}-B_{x,E}^{1/2}\|_{\rm op}.$$

Write $S=G_{x,E}^{1/2}$, $T=B_{x,E}^{1/2}$ and $Z=S-T$. Without a commutativity assumption,

$$SZ+ZT=G_{x,E}-B_{x,E}.$$

The unique solution is

$$Z=\int_0^\infty e^{-tS}(G_{x,E}-B_{x,E})e^{-tT}\,dt.$$

Differentiation of the integrand verifies the equation, and exponential decay proves convergence and uniqueness. Since $S\succeq\sqrt\gamma I$ and $T\succeq\sqrt{\gamma-\varepsilon}I$, the integral gives the first bound. Moreover,

$$\|C_x(I-P_x)\|_{\rm op}^2
=\|(I-P_x)C_x^*C_x(I-P_x)\|_{\rm op}\leq\varepsilon,$$

because $G_x$ vanishes on $E_x^\perp$. The triangle inequality gives the second bound.

## 2.2 Proof of the equivalences

An exact factor is injective on the range of $G_x$, so statement 1 implies statement 2. It also has zero error, proving statement 3.

Conversely let $T_x:E_x\to\mathbb C^m$ be a continuous bundle injection. Extend it to the ambient input by $T_xP_x$. Its Gramian is positive definite on $E_x$; compactness supplies a uniform positive lower bound. Its polar normalization

$$V_x=T_x(T_x^*T_x)_{E}^{-1/2}$$

is an isometric embedding, and $V_xG_x^{1/2}$ is an exact factor. Finally, statement 3 implies statement 1 by the repair formula. This proves the theorem.

**Consequences.** The minimum continuous exact row count depends only on the range bundle, not on its positive eigenvalue magnitudes. Approximation strictly below the positive spectral gap cannot reduce that row count. At the gap itself this implication can fail: for $G=\operatorname{diag}(\gamma,0,\ldots,0)$, the zero factor has error exactly $\gamma$ and loses the entire positive direction. The strict inequality is essential. When $G=0$ identically, a zero-row factor is exact; the positive-gap statement is unnecessary.

The polar normalization and square-root estimates use established matrix analysis. The contribution here is their explicit integration into the continuous-width problem, with all domain restrictions and correction costs specified.

\newpage

# 3. The sharp Grassmannian minimax law

**Theorem 2 (global width and complete directional failure).** Fix $1\leq r<n$ and $\lambda>\mu\geq0$. Over the entire complex Grassmannian define

$$G_P=\lambda P+\mu(I-P).$$

For every continuous $C:\operatorname{Gr}(r,n)\to\mathbb C^{m\times n}$ with $m<n$, there are a parameter $P_0$ and a unit vector $v_0\in\operatorname{ran}P_0$ such that

$$C(P_0)v_0=0.$$

Therefore

$$\boxed{\inf_{C\ {
m continuous}}\sup_P
\|C(P)^*C(P)-G_P\|_{\rm op}
=\begin{cases}\lambda,&0\leq m<n,\\0,&m\geq n.\end{cases}}$$

For $\mu=0$, the exact continuous row width is $n$, although every target Gramian has rank $r$. For comparison, independent pointwise optimization gives

$$\inf_{C\in\mathbb C^{m\times n}}
\|C^*C-G_P\|_{\rm op}
=\begin{cases}\lambda,&m<r,\\\mu,&r\leq m<n,\\0,&m\geq n.\end{cases}$$

## 3.1 The classical topological input

We use two standard facts, stated explicitly. Complex projective space has integral cohomology ring

$$H^*(\mathbb{CP}^{N};\mathbb Z)=\mathbb Z[h]/(h^{N+1}),\qquad\deg h=2.$$

For its tautological line bundle $L$, choose the sign convention $c(L)=1-h$. Total Chern classes satisfy the Whitney product identity $c(F\oplus H)=c(F)c(H)$; a rank-$q$ bundle has $c_j=0$ for $j>q$; a trivial bundle has total class one. These are classical results [M5, M8]. They supply the obstruction below; no new characteristic-class construction is claimed.

## 3.2 Proof of the forced kernel

If $m<r$, every restriction $C(P)|_{\operatorname{ran}P}$ has a kernel by rank-nullity. Suppose instead $r\leq m<n$ and, towards a contradiction, the restriction is injective at every $P$.

Fix an $(r-1)$-plane $E_0\subset\mathbb C^n$. Its orthogonal complement has dimension $n-r+1$. Vary a line $L\subset E_0^\perp$ and restrict the parameter to

$$E_P=E_0\oplus L,\qquad [L]\in\mathbb{CP}^{n-r}.$$

The restricted tautological rank-$r$ bundle is $\underline{\mathbb C}^{r-1}\oplus L$. The assumed injection embeds it into the trivial rank-$m$ output bundle. Take its orthogonal-complement bundle $Q$, of rank $m-r$. The Whitney identity gives

$$1=c(\underline{\mathbb C}^{r-1}\oplus L)c(Q)
=(1-h)c(Q),$$

and hence, in the truncated cohomology ring,

$$c(Q)=1+h+h^2+\cdots+h^{n-r}.$$

The coefficient $c_{m-r+1}(Q)=h^{m-r+1}$ is nonzero because $m-r+1\leq n-r$. It must also vanish because $Q$ has rank $m-r$. This contradiction proves the forced kernel. The argument works for every intermediate rank, including $r=n-1$.

## 3.3 Proof of sharpness and pointwise comparison

At the forced unit vector, $v_0^*G_{P_0}v_0=\lambda$ and $\|C(P_0)v_0\|=0$. Thus the operator error is at least $\lambda$. The identically zero factor attains error $\lambda$, proving the minimax value for $m<n$.

For $m\geq n$, the continuous square-root factor

$$C(P)=\sqrt\lambda\,P+\sqrt\mu\,(I-P)$$

is exact; append zero rows if required. The pointwise formula follows by retaining the largest $m$ eigenvalues, or by the kernel argument in Proposition A12. This completes the proof.

The theorem is stronger than a statement that an exact global frame cannot be selected. It shows that every smaller continuous encoder loses an entire target direction somewhere, even when the background eigenvalue is positive. Compressing its Gramian to $\operatorname{ran}P$ and applying Theorem 1 gives another formulation: error below $\lambda$ would repair $C(P)P$ into a forbidden embedding of the tautological bundle.

# 4. Labelled completions and their exact global cost

**Theorem 3 (arbitrary-rank labelled reserve law).** Let $k\geq1$, $\alpha_j>0$, and $a=\sum_j\alpha_j\leq1$. For $P\in\operatorname{Gr}(r,n)$ put

$$G_j(P)=\alpha_jP,\qquad
V(P)=I+(\sqrt{1-a}-1)P.$$

Then $V(P)^*V(P)+\sum_jG_j(P)=I$. At one fixed parameter the minimum separated reserve width is $kr$ and the minimum pooled width is $r$. For continuous factors in fixed coordinates these widths are exactly $kn$ and $n$.

If label $j$ is allowed operator error at most $\varepsilon_j$, its minimum continuous width is $n$ when $\varepsilon_j<\alpha_j$, and zero when $\varepsilon_j\geq\alpha_j$. The separated minima add. A pooled error threshold $\varepsilon$ gives width $n$ for $\varepsilon<a$, and zero for $\varepsilon\geq a$.

**Proof.** The ledger follows by squaring $V$ on $\operatorname{ran}P$ and its complement. Pointwise Gramian factorization gives the rank minima. Apply Theorem 2 with $\lambda=\alpha_j$, $\mu=0$ independently to each orthogonal labelled output space, and with $\lambda=a$ to the pool. The continuous factors $D_j(P)=\sqrt{\alpha_j}P$ attain all exact bounds; zero factors attain the permitted full-weight error. Hence the error-threshold conclusions are also sharp.

The complete encoder and decoder are

$$J(P)x=(V(P)x,D_1(P)x,\ldots,D_k(P)x),$$

$$J(P)^*J(P)=I,\qquad
x=V(P)^*x_{\rm vis}+\sum_jD_j(P)^*m_j.$$

The row cost arises from the chosen global coordinate requirement, not a failure of reversibility.

## 4.1 Realization by an ordered attenuation word

Let $\rho_0=1$ and $\rho_j=1-\sum_{i\leq j}\alpha_i$. Define

$$A_j(P)=I+(\sqrt{\rho_j/\rho_{j-1}}-1)P.$$

All denominators are positive: only the final $\rho_j$ may be zero. Earlier attenuation scales each support direction by $\sqrt{\rho_{j-1}}$. The $j$th event therefore transfers the chronological weight

$$\rho_{j-1}(1-\rho_j/\rho_{j-1})P=\alpha_jP.$$

At any fixed parameter each rank-$r$ event is a block of $r$ commuting rank-one events along an orthonormal support basis. The visible block operator is independent of that basis. The transfer factors need a basis choice locally, or $n$ fixed output coordinates globally. The construction does not assume a forbidden globally continuous support basis.

# 5. Reference frames and the exact failure radius

Fix a reference frame $W\in\mathbb C^{n\times r}$, $W^*W=I_r$, and write $S=\operatorname{ran}W$. Define

$$M_W(P)=W^*PW,\qquad
\delta_W(P)=\sigma_{\min}(PW)=\sqrt{\lambda_{\min}(M_W(P))}.$$

The recognized domain is $D_W=\{P:\delta_W(P)>0\}$; its complement $Z_W$ is the blind locus. The margin is the cosine of the largest principal angle between the target subspace and the reference subspace.

**Theorem 4 (canonical factor and sharp distance to blindness).** On $D_W$,

$$F_W(P)=PW\,M_W(P)^{-1/2}$$

is a continuous orthonormal frame of $\operatorname{ran}P$. It is uniquely characterized by $F^*F=I_r$, $FF^*=P$, and $W^*F\succ0$. For $\lambda>0$,

$$C_W(P)=\sqrt\lambda\,F_W(P)^*$$

is an exact $r$-row factor of $\lambda P$, and is pointwise minimal. More generally $F_W(P)^*G_P^{1/2}$ factors any positive semidefinite $G_P$ with range $\operatorname{ran}P$.

For every $P\in\operatorname{Gr}(r,n)$,

$$\boxed{\inf_{Q\in Z_W}\|P-Q\|_{\rm op}=\delta_W(P).}$$

In particular $\|P-Q\|_{\rm op}<\delta_W(P)$ preserves recognition, and

$$\delta_W(Q)\geq\delta_W(P)-\|P-Q\|_{\rm op}.$$

## 5.1 Proof of the frame formula

The positive definiteness of $M_W$ gives $F^*F=I_r$. Its range is the range of $PW$, which is the whole rank-$r$ support, so $FF^*=P$. Also $W^*F=M_W^{1/2}\succ0$. For any other support frame $F'$, if $W^*F'$ is positive definite then

$$ (W^*F')(W^*F')^*=W^*PW=M_W.$$

The unique positive square root gives $W^*F'=M_W^{1/2}$, fixing the frame. The factor identities and rank minimum follow immediately. Continuity holds by positive-definite functional calculus.

## 5.2 Proof of the distance formula

If $Q\in Z_W$, some unit $w\in S$ has $Qw=0$. Therefore

$$\|P-Q\|_{\rm op}\geq\|(P-Q)w\|=\|Pw\|\geq\delta_W(P).$$

For the matching upper bound first suppose $0<\delta<1$. Choose a unit eigenvector $w\in S$ of $(WW^*)P|_S$ with eigenvalue $\delta^2$. Put

$$e=Pw/\delta,\qquad
v=(w-\delta e)/\sqrt{1-\delta^2}.$$

Here $e\in\operatorname{ran}P$, $v\in\ker P$, and $e,v$ are orthonormal. Their reference projections are $(WW^*)e=\delta w$ and $(WW^*)v=\sqrt{1-\delta^2}\,w$. Thus

$$z=\sqrt{1-\delta^2}\,e-\delta v$$

is a unit vector orthogonal to $S$. Replace the support direction $e$ by $z$:

$$Q=P-ee^*+zz^*.$$

This is a rank-$r$ orthogonal projector and contains $z\in S^\perp$, so $Q$ is blind. The replacement acts only on $\operatorname{span}\{e,v\}$ and its projector distance is $\sqrt{1-|e^*z|^2}=\delta$.

If $\delta=0$, $P$ is already blind. If $\delta=1$, the target equals $S$; replacing any unit support vector by a unit vector in $S^\perp$ gives a blind projector at distance one. Such a complement exists because $r<n$. This proves the exact distance. Finally, the least singular value changes by at most $\|(P-Q)W\|_{\rm op}\leq\|P-Q\|_{\rm op}$, proving the robustness inequality.

# 6. Conditioning of the canonical frame

**Theorem 5 (polar stability and necessary inverse-margin growth).** For two recognized projectors $P,Q$,

$$\boxed{\|F_W(P)-F_W(Q)\|_F
\leq\frac{2\|(P-Q)W\|_F}{\delta_W(P)+\delta_W(Q)}.}$$

Consequently, if $\eta=\|P-Q\|_{\rm op}<\delta_W(P)=\delta$, then

$$\|F_W(P)-F_W(Q)\|_F
\leq\frac{2\sqrt r\,\eta}{2\delta-\eta}.$$

The order $1/\delta$ cannot be replaced by a uniform constant as the blind locus is approached. These are quantitative stability bounds; the global best Frobenius constant for this projector-restricted family is not asserted.

**Proof.** We first prove a full-column-rank polar inequality. Write $A=UH$, $B=VK$ with $U^*U=V^*V=I_r$, $H,K\succ0$, and least singular values $a,b$. For $\Delta=U-V$,

$$\begin{aligned}
\operatorname{Re}\operatorname{tr}\Delta^*(A-B)
&=\operatorname{tr}\bigl[(I-\operatorname{Re}V^*U)H\bigr]
+\operatorname{tr}\bigl[(I-\operatorname{Re}U^*V)K\bigr]\\
&\geq\frac{a+b}{2}\|U-V\|_F^2.
\end{aligned}$$

Both Hermitian matrices in parentheses are positive semidefinite because $U^*V$ is a contraction. Frobenius Cauchy–Schwarz bounds the left side by $\|U-V\|_F\|A-B\|_F$. Cancel the nonzero factor, or use the trivial zero case, to obtain $\|U-V\|_F\leq2\|A-B\|_F/(a+b)$. Set $A=PW$, $B=QW$ to get the theorem. The second bound uses Theorem 4 and $\|(P-Q)W\|_F\leq\sqrt r\eta$.

For necessity of inverse-margin growth, keep $r-1$ reference vectors fixed and use orthonormal vectors $w\in S$ and $v\in S^\perp$ for the final support direction

$$u_\phi=\sqrt{1-\delta^2}\,v+\delta e^{i\phi}w.$$

The corresponding final canonical frame column is

$$f_\phi=e^{-i\phi}\sqrt{1-\delta^2}\,v+\delta w.$$

For a nonzero sufficiently small phase change,

$$\frac{\|F_\phi-F_0\|_{\rm op}}{\|P_\phi-P_0\|_{\rm op}}
=\frac1\delta,\qquad
\frac{\|F_\phi-F_0\|_F}{\|P_\phi-P_0\|_F}
=\frac1{\sqrt2\,\delta}.$$

These exact ratios prove the order claim. Multiplying the frame adjoint by $\sqrt\lambda$ multiplies all factor-displacement bounds by $\sqrt\lambda$.

# 7. The exact minimum reference atlas

An atlas of fixed rank-$r$ references is a list $W_1,\ldots,W_q$, with $W_j^*W_j=I_r$, whose recognized domains cover $\operatorname{Gr}(r,n)$. Equivalently every target frame $U$ has $\det(W_j^*U)\ne0$ for at least one $j$.

**Theorem 6 (sharp chart count with an explicit construction).** Over the complex Grassmannian,

$$\boxed{q_{\min}=r(n-r)+1.}$$

Choose any $q_{\min}$ distinct real numbers $t_1,\ldots,t_q$. Define the derivative-evaluation matrix $M(t)\in\mathbb R^{n\times r}$ by

$$M(t)_{jk}=\begin{cases}
\dfrac{j!}{(j-k)!}t^{j-k},&j\geq k,\\
0,&j<k,
\end{cases}
\qquad 0\leq j<n,\quad0\leq k<r,$$

and normalize it:

$$W(t)=M(t)\bigl(M(t)^*M(t)\bigr)^{-1/2}.$$

Then $W(t_1),\ldots,W(t_q)$ is a covering atlas, and no atlas with fewer fixed rank-$r$ references covers the whole space.

## 7.1 Lower bound by projective intersection

We state the classical algebraic input [M9]. A nonempty projective variety of complex dimension $d$ cannot avoid the common zero set of $q\leq d$ homogeneous linear equations. To see its relation to the standard height theorem, its affine cone has dimension $d+1$. Imposing $q$ equations leaves a component of dimension at least $d+1-q\geq1$; the common zero set cannot consist only of the cone origin. A nonzero point descends to a projective intersection.

The Plücker embedding realizes $\operatorname{Gr}(r,n)$ as a projective variety of dimension $d=r(n-r)$ in $\mathbb P(\bigwedge^r\mathbb C^n)$. A target frame $U=(u_1,\ldots,u_r)$ represents the line of $u_1\wedge\cdots\wedge u_r$. Each equation

$$\det(W_j^*U)=0$$

is a homogeneous linear equation in those exterior coordinates. Thus $q\leq d$ blind loci have a common target. The references fail to cover, giving $q\geq d+1$. The determinant's conjugated coefficients are fixed constants and do not affect the complex algebraic argument.

## 7.2 Upper bound by an exact polynomial argument

The first $r$ rows of $M(t)$ are triangular with diagonal $0!,1!,\ldots,(r-1)!$, so $M(t)$ has full column rank for every real $t$. Let $U\in\mathbb C^{n\times r}$ be any orthonormal target frame and associate its columns with the independent polynomials

$$f_i(t)=\sum_{j=0}^{n-1}U_{ji}t^j.$$

The matrix $M(t)^*U$ consists of their derivatives of orders zero through $r-1$. Its determinant is their Wronskian

$$\mathcal W_U(t)=\det\bigl(f_i^{(k)}(t)\bigr)_{0\leq k<r,\ 1\leq i\leq r}.$$

An invertible change of polynomial basis multiplies the determinant by a nonzero constant. Use elimination to obtain basis polynomials with distinct leading degrees $0\leq k_1<\cdots<k_r<n$ and nonzero leading coefficients $a_i$. The highest coefficient of the Wronskian is

$$\left(\prod_i a_i\right)\prod_{i<j}(k_j-k_i)\ne0,$$

at degree $\sum_i k_i-r(r-1)/2\leq r(n-r)$. The coefficient identity follows because the determinant of the falling factorials $(k_i)_k$ is the Vandermonde determinant: these are monic polynomials of successive degrees in $k_i$. Hence the Wronskian is a nonzero polynomial of degree at most $d$.

It cannot vanish at $d+1$ distinct real nodes. At a node where it is nonzero,

$$\det(W(t)^*U)
=\det\bigl(M(t)^*M(t)\bigr)^{-1/2}\mathcal W_U(t)\ne0.$$

Thus the constructed references cover every target and attain the lower bound. This proves the theorem.

This construction is an application of the classical Wronski map [M10]. The optimal chart count and the construction are used here as exact resources in the continuous-factorization problem. The statement is for fixed linear reference frames over the complex family; a restricted real parameter family or more general chart types require separate analysis.

## 7.3 A sharp four-chart failure in $\operatorname{Gr}(2,4)$

Here $d=4$ and the minimum is five. Let

$$a=\frac{5+\sqrt{73}}6,\qquad b=\frac{-5+\sqrt{73}}2,$$

and take the two-plane of coefficient vectors of $f(t)=t^2-a$ and $g(t)=t^3-bt$. Its Wronskian is

$$f(t)g'(t)-f'(t)g(t)=t^4-5t^2+4=(t^2-1)(t^2-4).$$

Consequently the four distinct references at $t=-2,-1,1,2$ are all blind to this one two-plane. The additional node $t=0$ detects it. The polynomial calculation gives an explicit failure witness, rather than inferring failure from finite sampling.

# 8. A reference atlas with a conditioning certificate

Minimum cardinality and optimal conditioning are different design objectives. The Wronskian atlas covers the compact family, so each fixed node set has a positive worst-case margin. Theorem 6 does not give its optimal numerical value. A larger coordinate atlas gives a simple explicit uniform bound.

**Theorem 7 (coordinate coverage and guaranteed margin).** For each $r$-element index set $I\subset\{1,\ldots,n\}$, let $W_I$ consist of the corresponding standard basis columns. With $N=\binom nr$,

$$\boxed{\max_{|I|=r}\delta_{W_I}(P)\geq\binom nr^{-1/2}
\quad\text{for every }P.}$$

Thus a maximum-margin coordinate selector provides an exact local $r$-row factor whose failure radius is at least $N^{-1/2}$. The bound is a certificate, not an asserted optimum for all $r$.

**Proof.** If $P=UU^*$ and $U^*U=I_r$, Cauchy–Binet gives

$$\sum_{|I|=r}|\det(W_I^*U)|^2=\det(U^*U)=1.$$

Some determinant has magnitude at least $N^{-1/2}$. All singular values of $W_I^*U$ are at most one, so their product is at most their smallest member. Its smallest member equals $\sigma_{\min}(PW_I)$. Hence that chart has margin at least $N^{-1/2}$, and so does a maximum-margin selector.

| $n$ | $r$ | Pointwise rows | Continuous global rows | Minimum fixed references | Coordinate references | Certified coordinate margin |
|--:|--:|--:|--:|--:|--:|--:|
| 4 | 2 | 2 | 4 | 5 | 6 | $1/\sqrt6$ |
| 6 | 3 | 3 | 6 | 10 | 20 | $1/\sqrt{20}$ |
| 8 | 3 | 3 | 8 | 16 | 56 | $1/\sqrt{56}$ |
| 8 | 4 | 4 | 8 | 17 | 70 | $1/\sqrt{70}$ |

For rank one, a minimal coordinate atlas is an orthonormal basis. Proposition B6 proves its stronger optimality statement: it uniquely maximizes the worst-case squared overlap among $n$ unit references. No corresponding higher-rank conditioning optimum is asserted here.

![Exact continuous minimax error and the two reference-atlas counts. These are formulas from Theorems 2, 6 and 7, not experimental fits.](FIGURE_01.png){width=95%}

## 8.1 A complete local encoding protocol

Precompute either the $d+1$ derivative-evaluation frames or the $N$ coordinate frames. At a target projector $P$, evaluate each $\delta_j=\sigma_{\min}(PW_j)$, choose a largest-margin chart, and compute $F_j(P)$ by a thin polar decomposition. Return the label-specific amplitudes $m_\ell=\sqrt{\alpha_\ell}F_j(P)^*x$ with the visible state $V(P)x$. Store the chart index with the output.

On an overlap, two frames satisfy

$$F_j(P)=F_i(P)Q_{ij}(P),\qquad Q_{ij}(P)=F_i(P)^*F_j(P)\in U(r).$$

To change charts without changing the represented vector, transform reserve coordinates by $m_j=Q_{ij}(P)^*m_i$. The transition matrices obey $Q_{ij}Q_{jk}=Q_{ik}$. The local decoder uses the frame associated with the stored index. Neither routing nor reference storage is included in the count of reserve amplitudes.

# 9. Canonical comparison and noncommuting holonomy

**Theorem 8 (higher-rank transport and graph compatibility).** Let $P,Q$ have equal positive rank $r$ and suppose $QPQ$ is positive definite on $\operatorname{ran}Q$. Define

$$T(P\leftarrow Q)=PQ\,(QPQ)_{\operatorname{ran}Q}^{-1/2}.$$

This is a continuous canonical partial isometry on the transverse-pair domain, with

$$T^*T=Q,\qquad TT^*=P,\qquad
T(Q\leftarrow P)=T(P\leftarrow Q)^*.$$

It is covariant under a common unitary change of coordinates. On a finite connected graph of such comparisons, compatible orthonormal frames at all vertices exist if and only if every closed-cycle transport is the identity on its initial subspace. A cycle matrix in an initial frame belongs to $U(r)$ and changes by conjugation when that frame changes. For $r\geq2$, cycle matrices can fail to commute.

**Proof.** The positive-definite restriction makes $PQ:\operatorname{ran}Q\to\operatorname{ran}P$ injective and hence bijective. Its polar normalization gives the displayed formula and $T^*T=Q$. Its full rank onto the target gives $TT^*=P$. The adjoint is the polar partial isometry of $QP$, giving the reverse relation. Functional calculus commutes with unitary conjugation, proving covariance; the positive gap on each local pair domain proves continuity.

If edge transport carries every source frame to the target frame, compositions around cycles return the initial frame, so all cycle operators are the identity. Conversely choose a root frame and transport it along a spanning tree. These choices are unambiguous on the tree. Every additional edge closes a cycle; an identity cycle makes that edge compatible as well. This proves the criterion. Expressing a cycle operator $H$ in a new initial frame $FQ_0$ changes its matrix from $F^*HF$ to $Q_0^*(F^*HF)Q_0$.

## 9.1 An exact noncommuting example

Work in $\mathbb C^3$ at rank two, with initial frame $F_0=(e_1,e_2)$. Let $t=3/5$ and use the target planes with frames

$$F_A=\left(\frac{e_1+te_3}{\sqrt{1+t^2}},e_2\right),\quad
F_B=\left(e_1,\frac{e_2+te_3}{\sqrt{1+t^2}}\right),\quad
F_C=\left(e_1,\frac{e_2+ite_3}{\sqrt{1+t^2}}\right).$$

All relevant pair comparisons are invertible. The cycles $0\leftarrow A\leftarrow B\leftarrow0$ and $0\leftarrow A\leftarrow C\leftarrow0$ have matrices

$$H_R=\begin{pmatrix}a&b\\-b&a\end{pmatrix},\qquad
H_C=\begin{pmatrix}a&ib\\ib&a\end{pmatrix},$$

where $a=10\sqrt{34}/59$, $b=9/59$, and $a^2+b^2=1$. Indeed the intermediate overlap matrices are triangular with diagonal $1/\sqrt{1+t^2}$ and upper entry respectively $t^2/(1+t^2)$ and $it^2/(1+t^2)$; their polar factors are the matrices above. Therefore

$$H_RH_C-H_CH_R=2ib^2\operatorname{diag}(1,-1),$$

$$\boxed{\|H_RH_C-H_CH_R\|_{\rm op}=\frac{162}{3481}>0.}$$

Order of comparison histories can therefore affect the transported frame even when both histories return to the same target projector. This higher-rank effect requires matrix-valued transition data; scalar phase alone cannot represent it. The construction is consistent with the classical geometric transport context [M6, M7].

# 10. Exact repair and stable decoding

**Theorem 9 (repaired ledger and noisy decoder).** Suppose a continuous visible map $V_x$ and positive semidefinite label Gramians satisfy

$$V_x^*V_x+\sum_{j=1}^kG_{j,x}=I.$$

Assume each $G_j$ has positive constant rank with uniform positive gap $\gamma_j$. If continuous factors $C_j$ satisfy

$$\sup_x\|C_{j,x}^*C_{j,x}-G_{j,x}\|_{\rm op}
\leq\varepsilon_j<\gamma_j,$$

then Theorem 1 repairs them, without extra rows, to $D_j$. The complete map $J_x=(V_x,D_{1,x},\ldots,D_{k,x})$ is an isometry. Its adjoint decodes every input exactly and amplifies no additive output noise:

$$\|J_x^*(J_xv+e)-v\|\leq\|e\|.$$

Even without repair, let $K_x=(V_x,C_{1,x},\ldots,C_{k,x})$ and $\eta=\sum_j\varepsilon_j<1$. Then

$$ (1-\eta)I\preceq K_x^*K_x\preceq(1+\eta)I,$$

and its continuous least-squares left decoder

$$L_x=(K_x^*K_x)^{-1}K_x^*$$

satisfies

$$L_xK_x=I,\qquad
\boxed{\|L_x\|_{\rm op}\leq(1-\eta)^{-1/2}.}$$

This final noise-amplification bound is sharp given only the error budget.

**Proof.** Repair gives $D_j^*D_j=G_j$, so the original ledger becomes $J_x^*J_x=I$. Thus $J_x^*$ has norm one. For the unrepaired map,

$$K_x^*K_x-I=\sum_j(C_{j,x}^*C_{j,x}-G_{j,x}),$$

whose operator norm is at most $\eta$. The positive lower bound gives invertibility and continuity of the decoder. Its singular values are the reciprocals of the nonzero singular values of $K_x$, giving the bound. For sharpness take a one-dimensional input, $V=0$, $G_1=1$, and $C_1=\sqrt{1-\eta}$. Its Gram error is $\eta$ and the decoder norm is exactly $(1-\eta)^{-1/2}$. At $\eta=1$ the zero factor is permitted and decoding can fail.

The hypotheses themselves are resource constraints. For the full Grassmannian family, Theorem 2 prevents satisfying a sub-gap tolerance with too few continuous reserve rows. Decoding estimates cannot remove that topological obstruction.

# 11. Reproducibility and verification

The package contains executable implementations for finite-word Gramians, rank-one geometry, arbitrary-rank factor repair, reference frames, both atlases, and canonical subspace transport. Run `python3 verify.py` after installing `requirements.txt`. The checker writes `VERIFICATION.json`. `python3 verify_continuous.py` runs the new arbitrary-rank audit independently.

The proofs establish universal statements. Computation checks exact examples, independent algebraic identities, constructive witnesses, decoder behaviour, and numerical conditioning. In particular, sampling does not establish the topological lower bounds in Theorem 2 or the projective lower bound in Theorem 6.

## 11.1 Exact arithmetic

The previous finite-word audit checks 6,270 rational words and 18,738 prefixes. The new independent Wronskian audit checks all 494 coordinate polynomial subspaces for $2\leq n\leq8$, evaluating 6,124 jet determinants with rational arithmetic. It also checks 64 independent integer polynomial subspaces, computing determinant polynomials directly, verifying their degree bound, and testing their values at $d+1$ distinct nodes. These are actual executed counts.

## 11.2 Arbitrary-rank numerical audit

With seed 30001004, dimensions $2\leq n\leq8$, and every $1\leq r<n$, the new suite performs 224 cases each for factor repair, canonical reference factors, inside-radius perturbations, Wronskian atlas selection, coordinate atlas selection, comparison cycles, and noisy decoding. It checks 252 constructed nearest-blind projectors, 140 exact inverse-margin ratios, 28 gap-boundary rejections, and 28 explicit blind scanner families. It independently tests the four-reference $\operatorname{Gr}(2,4)$ failure and the noncommuting $U(2)$ example.

The repaired Gramian residual is at most $5.50\times10^{-15}$; the nearest-blind distance residual is at most $7.78\times10^{-16}$; the arbitrary-rank decoder residual is at most $9.14\times10^{-15}$. Canonical transport residuals reach $2.72\times10^{-13}$ for less well-conditioned sampled comparisons. These are finite floating-point residuals, not certified uniform error bounds. The numerical rank tolerance is $10^{-12}$ where declared; exact rank is defined algebraically in the proofs.

The smallest maximum Wronskian-atlas margin seen in these 224 sampled targets is approximately $0.23018$. It is a sample statistic for the tested dimensions and node sets. The coordinate-atlas bound in Theorem 7 is the proved uniform certificate.

## 11.3 Scope and completion record

| Item | Actual status |
|:--|:--|
| Core theorem statements and proofs | 9 of 9 written and internally checked |
| Supporting finite-word and rank-one results | 29 propositions retained with proofs |
| New exact Wronskian subspace cases | 558 executed |
| Reproducible check programs | Executed; full counts in `VERIFICATION.json` |
| PDF and release package | Built, inspected, and checksummed |
| Proof-assistant certification | Not performed |
| Independent specialist review | Not performed |
| Empirical application benchmark | Not performed |
| Worldwide priority and broader scientific importance | Not established |

Constant-rank finite-dimensional families, continuous fixed-coordinate complex-linear factors, fixed linear references, and the stated metrics are essential assumptions. Changing-rank targets, arbitrary adaptive event policies, infinite-dimensional systems, and optimized higher-rank atlas conditioning remain separate problems.

# 12. Mathematical provenance and research contribution

The strongest contribution in this release is the explicit connection among three exact resources: the embedding dimension of a continuous range bundle, the positive-gap threshold for repairable approximation, and the minimum atlas of fixed reference frames. This connects a sharp impossibility theorem with implementable factors, certified failure radii, transition data, and stable decoders.

The range-bundle formulation and Chern-class obstruction draw on classical vector-bundle theory. The chart lower bound uses classical projective intersection; the construction uses the established Wronski map. Gramian factorization, polar decomposition, Cauchy–Binet, geometric phases, and projection transport also have established provenance. Their inputs are named where used and listed below. The calculations, formulations, and proofs supplied here define a rigorous research synthesis; they do not certify that its precise statements have never appeared elsewhere. The targeted prior-art review is included in `NOVELTY_REVIEW.md`.

Potential applications concern continuous signal routing, parameter-dependent reduced representations, and reversible labelled accounting. Their relevance depends on a system actually imposing the matrix continuity and information requirements used here. No application performance claim is inferred from the theorem alone.

# Appendix A. Finite-word chronological accounting

The event word and its matrices are fixed and known, all reserves start empty, and every stated identity holds for arbitrary complex input vectors. Prefix compression is asserted on the reachable subspace generated by those conditions. These assumptions distinguish the chronological accounting problem from unrestricted devices with arbitrary preloaded memories.

## A.1. Finite-word definitions

Fix an integer $n\geq 1$ and the Hilbert space $V=\mathbb C^n$ with inner product $\langle x,y\rangle=x^*y$. The star denotes conjugate transpose. A Hermitian matrix $G$ is positive semidefinite, written $G\succeq0$, when $x^*Gx\geq0$ for every $x$. Its range and kernel are $\operatorname{ran}G$ and $\ker G$. All dimensions and linear maps are over $\mathbb C$, except where a real subfamily is explicitly stated.

An **attenuation event** is a tuple

$$e_t=(u_t,c_t,\ell_t,U_t),$$

where $\|u_t\|=1$, $0\leq c_t\leq1$, the label $\ell_t$ is either $\beta$ or $\gamma$, and $U_t$ is unitary. Set

$$s_t=\sqrt{1-c_t^2},\qquad
A_t=I+(c_t-1)u_tu_t^*,\qquad L_t=U_tA_t.$$

Thus $A_t$ multiplies the component along $u_t$ by $c_t$ and leaves the orthogonal complement unchanged. We may write $c_t=\cos\theta_t$ with $0\leq\theta_t\leq\pi/2$, but cosine notation is convenient for exact rational calculations. An event is **active** if $c_t<1$.

The visible state evolves by

$$x_t=L_tx_{t-1},\qquad P_0=I,\qquad P_t=L_t\cdots L_1,$$

so that $x_t=P_tx_0$. The uncompressed reserve output at event $t$ is the scalar

$$d_t=s_tu_t^*x_{t-1}=s_tu_t^*P_{t-1}x_0.$$

The same initial vector determines all outputs. Define the pulled-back normal and row

$$v_t=P_{t-1}^*u_t,\qquad b_t=s_tv_t^*.$$

For a prefix ending at $T$, stack the rows $b_t$ of each label into a matrix $S_\ell$. A label that never occurs has the zero-row matrix. Its **chronological reserve Gramian** is

$$G_\ell=S_\ell^*S_\ell
=\sum_{\substack{t\leq T\\\ell_t=\ell}}
(1-c_t^2)v_tv_t^*.$$

The chronological energy for that label is

$$E_\ell(x_0)=\|S_\ell x_0\|^2=x_0^*G_\ell x_0.$$

An exact separated reserve encoder consists of complex-linear maps

$$C_\beta:V\to\mathbb C^{m_\beta},\qquad
C_\gamma:V\to\mathbb C^{m_\gamma}$$

satisfying $\|C_\ell x\|^2=E_\ell(x)$ for every $x$. The two output spaces are orthogonal summands. Their dimensions count **persistent reserve coordinates**, not bytes, physical cells, qubits, or the number of scalar energy readouts.

Pooling permits one map $C:V\to\mathbb C^m$ satisfying $\|Cx\|^2=E_\beta(x)+E_\gamma(x)$. It removes the requirement for two orthogonal reserve outputs. It does not necessarily prevent separate scalar energy readouts; Proposition A3 makes that distinction precise.

## A.2. Chronological Gramian factorization

**Proposition A1 (exact minimum and causal completion).** For every fixed finite response word with initially empty reserves, let

$$R_\ell=\operatorname{ran}G_\ell,\qquad
r_\ell=\dim R_\ell,\qquad
G=G_\beta+G_\gamma.$$

Then the following statements hold.

1. The exact ledger is

   $$P_T^*P_T+G_\beta+G_\gamma=I.$$

2. The minimum separated width is

   $$m_{\rm sep}=r_\beta+r_\gamma.$$

3. The minimum pooled width is

   $$m_{\rm pool}=\operatorname{rank}G.$$

4. The exact excess required by orthogonal reserve labels is

   $$\boxed{m_{\rm sep}-m_{\rm pool}
   =\dim(R_\beta\cap R_\gamma).}$$

5. Minimum-width factors give an isometric completion

   $$J_Tx=(P_Tx,C_\beta x,C_\gamma x),\qquad J_T^*J_T=I.$$

   Every initial state is therefore recoverable from the complete output.

6. The minimum widths can be attained at **every prefix** by a causal update depending only on that prefix and the current state. The norm-preserving implementation is asserted on the reachable subspace generated from an arbitrary $x_0$ and empty reserves. It is not an unrestricted no-feedback unitary acting identically on arbitrary preloaded reserves.

### Proof of the ledger

Since $u_tu_t^*$ is an orthogonal projector,

$$A_t^*A_t=I-(1-c_t^2)u_tu_t^*.$$

Unitarity of $U_t$ gives $L_t^*L_t=A_t^*A_t$. Consequently

$$\begin{aligned}
P_{t-1}^*P_{t-1}-P_t^*P_t
&=P_{t-1}^*(I-L_t^*L_t)P_{t-1}\\
&=(1-c_t^2)v_tv_t^*=b_t^*b_t.
\end{aligned}$$

Summing over time telescopes to $I-P_T^*P_T$. Partitioning the summands by their labels proves the first statement. Every summand is positive semidefinite, so there is no cancellation hidden in this ledger.

### A factorization lemma and the minimum

We need a standard finite-dimensional fact, which we prove here.

**Lemma.** For $G\succeq0$, the minimum number of rows of a matrix $C$ satisfying $C^*C=G$ is $\operatorname{rank}G$.

**Proof.** If $C^*C=G$, then $\ker C=\ker G$, because $x^*Gx=\|Cx\|^2$. Hence $\operatorname{rank}C=\operatorname{rank}G$, and the row count cannot be smaller. Conversely diagonalize $G=V\Lambda V^*$, retaining only its $r$ positive eigenvalues. The matrix $C=\Lambda_+^{1/2}V_+^*$ has $r$ rows and satisfies $C^*C=G$. For $G=0$, the zero-row factor works. This proves the lemma.

Equality of the quadratic forms for every $x$ is equivalent to $C_\ell^*C_\ell=G_\ell$. Applying the lemma independently to the two orthogonal summands proves statement 2. Applying it to $G$ proves statement 3.

For positive semidefinite $G_\beta,G_\gamma$,

$$\ker(G_\beta+G_\gamma)=\ker G_\beta\cap\ker G_\gamma.$$

Taking orthogonal complements gives

$$\operatorname{ran}G=R_\beta+R_\gamma.$$

The dimension formula for two subspaces now gives statement 4. Combining the factors with the ledger gives $J_T^*J_T=I$, proving statement 5.

### Causal prefix-optimal compression

Suppose prefix $t-1$ is already encoded by factors $C_{\ell,t-1}$ and memories $m_{\ell,t-1}=C_{\ell,t-1}x_0$. At event $t$ only the label $\ell_t$ receives a new row. Form

$$B_t=\begin{pmatrix}C_{\ell_t,t-1}\\b_t\end{pmatrix}.$$

Then $B_t^*B_t=G_{\ell_t,t}$. Select a minimum-row factor $C_{\ell_t,t}$ of this Gramian. There is a column isometry $V_t$ with

$$B_t=V_tC_{\ell_t,t},\qquad V_t^*V_t=I.$$

To see this, define $V_t(C_{\ell_t,t}x)=B_tx$ on the factor's range. The two factors have the same kernel, so this definition is well-defined; their equal Gramians make it isometric. The minimum-row factor is surjective onto its output space, so this defines all columns of $V_t$.

The update is

$$m_{\ell_t,t}=V_t^*
\begin{pmatrix}m_{\ell_t,t-1}\\s_tu_t^*x_{t-1}\end{pmatrix},$$

while the other reserve is unchanged and $x_t=L_tx_{t-1}$. The stacked vector on the right lies in $\operatorname{ran}B_t$, so the compression loses neither its norm nor its information. This update uses stored memory, the current visible state, and prefix matrices; it does not reread $x_0$ or use a future event.

Both $J_{t-1}$ and $J_t$ are isometries from $V$. The map $J_{t-1}x\mapsto J_tx$ is therefore an isometry between their reachable subspaces. After padding the smaller ambient space with zero coordinates, it extends to a unitary by completing orthonormal bases. The reserve ranks never decrease, as shown in Proposition A5, so such padding uses only newly required persistent coordinates. The extension is unrestricted off the reachable subspace, where it need not preserve the no-feedback visible rule. This proves statement 6 and completes Proposition A1.

## A.3. Co-moving active span

**Proposition A2 (co-moving response span).** Define the transport-only word

$$Q_0=I,\qquad Q_t=U_t\cdots U_1,$$

and the co-moving response normal $a_t=Q_{t-1}^*u_t$. Let

$$W=\operatorname{span}\{a_t:t\leq T,\ c_t<1\}.$$

Then

$$\boxed{\operatorname{ran}G=W,
\qquad m_{\rm pool}=\dim W,
\qquad\ker G=W^\perp.}$$

In particular, for fixed $u_t,U_t$ and the same set of active events, changing their strengths within $0\leq c_t<1$ does not change pooled width, even when some events have $c_t=0$.

**Proof.** In the co-moving frame $y_t=Q_t^*x_t$, the local map is

$$B_t=I+(c_t-1)a_ta_t^*.$$

Thus $P_{t-1}=Q_{t-1}B_{t-1}\cdots B_1$ and

$$v_t=B_1\cdots B_{t-1}a_t.$$

Every active $B_j-I$ has range contained in $\mathbb Ca_j$. Therefore $v_t-a_t$ lies in the span of the earlier active $a_j$. An induction shows

$$\operatorname{span}\{v_t:c_t<1\}
=\operatorname{span}\{a_t:c_t<1\}.$$

For a positive weighted sum of rank-one projectors, the kernel is the common orthogonal complement of the generating vectors. Applying this to $G=\sum_{t:c_t<1}(1-c_t^2)v_tv_t^*$ proves the result. No inverse of an attenuation map was used; complete directional transfer is allowed.

**Interpretation.** Thousands of active events do not automatically require thousands of persistent coordinates. The relevant count is the number of independent response directions after undoing transport.

## A.4. Linear observation and energy readouts

**Proposition A3 (minimum scanner rank).** Let a linear scanner observe $y=Hx_0$. Both functions $E_\beta(x_0)$ and $E_\gamma(x_0)$ can be determined from $y$ for every input if and only if

$$\ker H\subseteq\ker G.$$

Equivalently $W\subseteq\operatorname{ran}H^*$. The minimum scanner rank is

$$\boxed{r_{\rm scan}=\operatorname{rank}G=m_{\rm pool}.}$$

The readouts may be taken to be positive semidefinite quadratic forms $E_\ell=y^*K_\ell y$.

**Proof.** If $z\in\ker H$, the scanner gives the same observation for $z$ and $0$. Determining both energies forces $z^*G_\ell z=0$. Positivity gives $G_\ell z=0$ for both labels, so $z\in\ker G$.

Conversely, if the kernel inclusion holds, the quadratic forms are constant along the scanner's fibres. Let $H^\dagger$ denote the Moore-Penrose pseudoinverse. Since $x-H^\dagger Hx\in\ker H\subseteq\ker G_\ell$,

$$G_\ell=H^*K_\ell H,\qquad
K_\ell=(H^\dagger)^*G_\ell H^\dagger\succeq0.$$

The kernel inclusion implies $\operatorname{rank}H\geq n-\dim\ker G$. An orthonormal row basis of $W$ attains equality.

The distinction is essential: two reported numbers do not require two independent amplitude stores. A scanner can determine both labelled energies from the shared active subspace. Proposition A1 requires separate orthogonal output spaces, a stronger implementation constraint.

## A.5. Decoding and prefix growth

**Proposition A4 (stable decoding and transcript recovery).** A minimum separated completion has the exact decoder

$$\boxed{x_0=P_T^*x_T+C_\beta^*m_\beta+C_\gamma^*m_\gamma.}$$

Each labelled raw transcript is also recoverable from its own compressed reserve:

$$S_\ell x_0=W_\ell m_\ell$$

for a fixed column isometry $W_\ell$ depending on the word.

**Proof.** The decoder is $J_T^*J_Tx_0=x_0$. For the transcript, $S_\ell^*S_\ell=C_\ell^*C_\ell$, so the same well-defined isometry argument used in Appendix A gives $S_\ell=W_\ell C_\ell$. The decoder has operator norm one on the augmented space; small additive output error is not amplified by the decoder.

This statement concerns the complete complex transcript for a known word. Recovering an unknown event sequence, its labels, or its transports is a different inverse problem.

**Proposition A5 (one-coordinate prefix growth).** For every label, its minimum reserve width is nondecreasing along prefixes. At any event, the total separated width increases by either zero or one. It increases precisely when the active $v_t$ lies outside the range of the receiving label's previous Gramian.

**Proof.** The receiving Gramian gains the positive term $(1-c_t^2)v_tv_t^*$. The range of a positive sum is the sum of its ranges. Thus its rank stays the same or increases by one; the other label is unchanged. If $c_t=1$, the summand is zero.

## A.6. An order-dependent rank example

**Proposition A6 (a strict chronological effect).** There are two three-event words with the same labelled event multiset, the same active geometric span, and different separated reserve widths. In one word, repeated use of a single raw $\beta$ direction gives a two-dimensional $\beta$ reserve.

**Proof by exact construction.** Work in $\mathbb R^2\subset\mathbb C^2$ with all transports equal to $I$. Let

$$a=(1,0)^T,\qquad b=(3/5,4/5)^T,\qquad c_t=3/5.$$

Compare the words

$$\mathcal A=(a,\beta),(b,\gamma),(a,\beta),\qquad
\mathcal B=(a,\beta),(a,\beta),(b,\gamma).$$

In word $\mathcal A$, the first $\beta$ pulled-back normal is $v_1=a$. The third is

$$v_3=A_aA_ba
=\begin{pmatrix}321/625\\-24/125\end{pmatrix}.$$

Its second coordinate is nonzero, so the $\beta$ Gramian has rank two. The $\gamma$ Gramian has rank one, since its single pulled-back normal is $(9/25,4/5)^T\ne0$. Hence $m_{\rm sep}=3$.

In word $\mathcal B$, both $\beta$ normals are multiples of $a$, so its $\beta$ rank is one. The $\gamma$ normal is $(27/125,4/5)^T\ne0$, so its $\gamma$ rank is one. Thus $m_{\rm sep}=2$.

Both words have $W=\operatorname{span}\{a,b\}=\mathbb C^2$ and $m_{\rm pool}=2$. Their separated excesses are respectively one and zero. The example disproves the shortcut of computing a labelled reserve's width from its raw normals alone. The products themselves need not be equal; this is an order comparison for equal event multisets, not an assertion of product-preserving rearrangement.

| Word | $\beta$ width | $\gamma$ width | Pooled width | Separated excess |
|:--|--:|--:|--:|--:|
| $\beta$-$a$, $\gamma$-$b$, $\beta$-$a$ | 2 | 1 | 2 | 1 |
| $\beta$-$a$, $\beta$-$a$, $\gamma$-$b$ | 1 | 1 | 2 | 0 |

## A.7. Bounds and general label coarsening

**Proposition A7 (sharp separated-width bound).** If $r=\dim W$, then

$$r\leq m_{\rm sep}\leq2r\leq2n.$$

Both extremes are attainable. The upper extreme can occur with $G_\beta$ and $G_\gamma$ positive definite on $W$.

**Proof.** Each range $R_\ell$ is a subspace of $W$, so its dimension is at most $r$. The lower bound follows from $R_\beta+R_\gamma=W$. For the lower extreme, assign every active event the same label.

For the upper extreme, take an orthonormal basis $e_1,\ldots,e_r$ of an $r$-dimensional subspace and a fixed $0<c<1$. With identity transports, first charge each basis direction into $\beta$ and then charge each into $\gamma$. The first block gives $G_\beta=(1-c^2)\Pi_W$, while the second gives $G_\gamma=c^2(1-c^2)\Pi_W$. Both have rank $r$, establishing equality.

**Proposition A8 (general colour coarsening).** For $p$ labels with Gramians $G_1,\ldots,G_p$, the exact separated excess is

$$\sum_{j=1}^p\dim R_j-\dim(R_1+\cdots+R_p).$$

It equals the kernel dimension of the addition map

$$\Sigma:\bigoplus_{j=1}^pR_j\longrightarrow R_1+\cdots+R_p,
\qquad (z_1,\ldots,z_p)\mapsto\sum_jz_j.$$

Merging a group of labels replaces the sum of their ranks by the rank of their Gramian sum and cannot increase required width.

**Proof.** The factorization lemma supplies the separated and merged minima. Positivity gives the range-sum identity. The addition map is surjective, and rank-nullity gives the displayed excess.

For three or more labels, pairwise intersections do not suffice. Three distinct lines in a plane may have zero pairwise intersections while their direct sum has dimension three and their ordinary sum has dimension two. Their excess is one. This prevents extending the two-colour intersection formula by a naive sum of pairwise overlaps.

## A.8. Unrestricted no-feedback obstruction

**Proposition A9 (fixed-memory no-feedback obstruction).** Suppose a square unitary on visible state and a fixed finite reserve has the block form

$$\mathcal U=\begin{pmatrix}A&0\\B&C\end{pmatrix}.$$

The zero upper-right block requires the visible output to be independent of every possible reserve input. Then $A$ must be unitary and $B=0$. A nonzero directional transfer cannot be implemented in this form.

**Proof.** The upper-left block of $\mathcal U\mathcal U^*=I$ gives $AA^*=I$. Since $A$ is square and finite-dimensional, it is unitary. The upper-left block of $\mathcal U^*\mathcal U=I$ then gives $A^*A+B^*B=I$, so $B=0$.

This obstruction is why Proposition A1 specifies initially empty reserves and their reachable subspace. Causal compression is possible there because reserve and visible coordinates are correlated by a common initial state.

For comparison, a single event can use a fresh empty scalar reserve through

$$\mathcal J_t=
\begin{pmatrix}
A_t&-s_tu_t\\s_tu_t^*&c_t
\end{pmatrix}.$$

On $u_t\oplus\mathbb C$, this is the real rotation matrix with entries $c_t,-s_t,s_t,c_t$; on $u_t^\perp$ it is the identity. It is therefore unitary. Multiplying its visible rows by $U_t$ incorporates transport. The nonzero upper-right block makes the distinction from the prohibited form explicit.

## A.9. Visible inverse and coordinate covariance

**Proposition A10 (visible-only reset and coordinate covariance).** An exact visible-only reset by a unitary $R$, satisfying $RP_T=I$ for all initial states, exists if and only if every event is inactive. In that case $P_T=Q_T$ and $R=Q_T^*$.

For any invertible coordinate charts $q_t=F_tx_t$, define

$$M_t=F_t^{-*}F_t^{-1},\qquad
\widetilde P_T=F_TP_TF_0^{-1},\qquad
\widetilde G_\ell=F_0^{-*}G_\ell F_0^{-1}.$$

Then

$$\widetilde P_T^*M_T\widetilde P_T
+\widetilde G_\beta+\widetilde G_\gamma=M_0.$$

All exact reserve and scanner widths, including the separated excess, are unchanged by these chart transformations.

**Proof.** A visible-only unitary reset implies $P_T$ is unitary and hence $G=0$. Proposition A2 then gives $W=0$. The first active event, if any, would contribute a nonzero normal to $W$, so no active event is possible. Conversely, without active events the product is $Q_T$, which has the stated inverse.

For covariance, substitute the definitions and cancel $F_T$ with $M_T$. The remaining equation is the original ledger conjugated by $F_0^{-1}$. Invertible congruence preserves rank. It maps each Gramian range by the same invertible map $F_0^{-*}$, so intersection dimension is preserved as well.

A pure sequence of coordinate changes has intrinsic operators $L_t=I$. Its chart transitions are $F_tF_{t-1}^{-1}$ and telescope. Returning to the same chart returns exactly to the starting representation. By contrast, a physical transport loop may have $Q_T\ne I$ even when the reserve is empty. This is a transport holonomy, not an energy charge.

If $P_T$ is invertible but not unitary, an algebraic visible inverse exists. That is not a norm-preserving visible-only reset. Its conditioning can be poor, whereas the augmented adjoint decoder in Proposition A4 has norm one.

## A.10. Directional scanner criterion

**Proposition A11 (local scanner criterion).** At an active event, suppose a scanner observes $y=Hx_{t-1}$. The exact complex reserve amplitude $s_tu_t^*x_{t-1}$ can be produced by a linear function of $y$ for every current state if and only if

$$u_t\in\operatorname{ran}H^*.$$

The same condition is necessary and sufficient merely to determine its squared magnitude for every current state.

**Proof.** A linear amplitude readout exists precisely when the row $s_tu_t^*$ belongs to the row space of $H$. Since $s_t>0$, this is the stated condition. For the energy-only claim, apply Proposition A3 to the rank-one Gramian $s_t^2u_tu_t^*$. If $c_t=1$, the output is zero and no sensing condition is required.

The theorem concerns complete knowledge over all possible input vectors. A scanner may succeed on a restricted known state family with less information; no such prior restriction is assumed here.

## A.11. Pointwise spectral approximation

Exact rank counts can be fragile. A tiny nonzero independent component raises exact width even when its energy is negligible. An approximate statement resolves this without pretending numerical tolerances are exact algebra.

**Proposition A12 (optimal uniform energy approximation).** Let the eigenvalues of $G\succeq0$ be $\lambda_1\geq\cdots\geq\lambda_n\geq0$. Among all factors $C$ with at most $k$ rows, the minimum worst-case quadratic energy error is

$$\boxed{\inf_C\sup_{\|x\|=1}
\left|x^*Gx-\|Cx\|^2\right|=\lambda_{k+1},}$$

where $\lambda_{n+1}=0$. Consequently the minimum width for error at most $\varepsilon$ is the number of eigenvalues strictly greater than $\varepsilon$. For independently prescribed errors on two separated reserves, their approximate widths add.

**Proof.** Truncate the spectral expansion to the largest $k$ eigenvalues and take its positive square-root factor. The remaining operator norm is $\lambda_{k+1}$, giving the upper bound. For any factor with at most $k$ rows, $\ker C$ meets the span of the top $k+1$ eigenvectors in a nonzero vector. Normalize such a vector $x$. Then $\|Cx\|=0$ and $x^*Gx\geq\lambda_{k+1}$, giving the lower bound. Apply this argument independently to each orthogonal reserve.

**Proposition A13 (threshold stability and exact-rank jumps).** If two Hermitian Gramians satisfy $\|G'-G\|\leq\delta$, their ordered eigenvalues differ by at most $\delta$. A threshold width at $\varepsilon$ is therefore unchanged whenever every eigenvalue of $G$ has distance greater than $\delta$ from $\varepsilon$. Exact width can jump under arbitrarily small perturbations.

**Proof.** The bound $-\delta I\preceq G'-G\preceq\delta I$ gives the ordered eigenvalue bound by the min-max characterization. If the original eigenvalues lie strictly outside the threshold's $\delta$-neighbourhood, none can cross it. For a jump, consider $G=\operatorname{diag}(1,0)$ and $G'=\operatorname{diag}(1,\eta)$ with any $\eta>0$. Their exact ranks are one and two, although $\|G'-G\|=\eta$ can be arbitrarily small.

Coordinate charts that are not unitary require the transported metric. Ordinary eigenvalues of a non-unitarily transformed Gramian should not be used as intrinsic approximate-width thresholds. Normalize back to the intrinsic Hilbert metric first.

## A.12. The two-chart line obstruction

So far, a word and its factors are fixed. A different question asks for factors that vary continuously over a family of response directions. Choosing one optimal factor at each point does not prove that the choices can be glued continuously.

Let $\mathbb {CP}^1$ be the set of complex lines $\ell\subset\mathbb C^2$. A line determines an orthogonal projector $\Pi_\ell=uu^*$, where $u$ is any unit vector in that line. Multiplying $u$ by a unit complex phase does not change the projector. Thus the projector represents a response ray without a preferred phase convention.

**Proposition A14 (sharp global scalar-channel obstruction).** Fix $\lambda>0$. Each Gramian $G(\ell)=\lambda\Pi_\ell$ has rank one and a one-row factor at every individual $\ell$. Nevertheless there is no continuous family of one-row factors

$$C(\ell)\in\mathbb C^{1\times2},\qquad
C(\ell)^*C(\ell)=\lambda\Pi_\ell$$

on all of $\mathbb {CP}^1$. The minimum number of **fixed complex output coordinates** for a globally continuous factor is exactly two, achieved by

$$C_{\rm global}(\ell)=\sqrt\lambda\,\Pi_\ell.$$

### Self-contained proof by two charts

A one-row factor would give a continuous unit vector

$$v(\ell)=C(\ell)^*/\sqrt\lambda\in\ell.$$

Consider the northern and southern coordinates

$$u_N(z)=\frac{(1,z)^T}{\sqrt{1+|z|^2}},\qquad
u_S(w)=\frac{(w,1)^T}{\sqrt{1+|w|^2}}.$$

They represent the same line when $w=1/z$. On the equator $z=e^{i\phi}$,

$$u_S(e^{-i\phi})=e^{-i\phi}u_N(e^{i\phi}).$$

A global unit vector choice would therefore give continuous unit phases $f_N,f_S$ on the two coordinate disks, with

$$v=f_N(z)u_N(z)=f_S(w)u_S(w).$$

On the equator these satisfy

$$f_N(e^{i\phi})=f_S(e^{-i\phi})e^{-i\phi}.$$

The boundary of a continuous disk-to-circle map has winding number zero. One way to see this is to contract its boundary radially to the image of the centre: $f(re^{i\phi})$, with $r$ decreasing from one to zero, is a homotopy to a constant loop. Winding is invariant under homotopy, additive under multiplication of circle-valued loops, and negated by reversing traversal. These elementary facts follow from lifting a loop's phase to a continuous real argument on the parameter interval; its endpoint difference is an integer multiple of $2\pi$.

The left side therefore has winding zero. The right side has winding zero minus one, which is minus one. This contradiction proves that a continuous one-row factor does not exist.

Two rows suffice because

$$C_{\rm global}^*C_{\rm global}
=\lambda\Pi_\ell^2=\lambda\Pi_\ell,$$

and $\ell\mapsto\Pi_\ell$ is continuous. This proves sharpness and Proposition A14.

### What the obstruction does and does not say

The obstruction is classical nontriviality of the tautological line bundle, expressed as a reserve implementation problem. It is not a newly discovered failure of the Hopf fibration to admit a section. The construction's contribution is to make it an operative restriction alongside the chronological memory theorem.

A one-dimensional **bundle-valued** reserve works globally: its fibre at $\ell$ can be the line $\ell$ itself. The impossibility concerns one globally fixed scalar coordinate. Alternatively, local scalar charts work with the transition phase $e^{-i\phi}$. Thus the options are a nontrivial reserve bundle, multiple local charts, or a redundant fixed-coordinate representation.

## A.13. Two labelled reserves over a line family

**Proposition A15 (local versus globally continuous labelled width).** Consider the following family of two-event words on $\mathbb C^2$, indexed by $\ell\in\mathbb {CP}^1$:

- both events use the same projector $\Pi_\ell$;
- both transports are $I$ and both cosines are $1/\sqrt2$;
- the first label is $\beta$ and the second is $\gamma$.

The event depends only on the projector, so no global vector phase is needed to define the visible maps. Its Gramians and final visible product are

$$G_\beta=\tfrac12\Pi_\ell,\qquad
G_\gamma=\tfrac14\Pi_\ell,\qquad
P_2=I-\tfrac12\Pi_\ell.$$

For each individual direction,

$$m_{\rm sep}^{\rm local}=2,\qquad m_{\rm pool}^{\rm local}=1.$$

For factors continuous over all directions with fixed output coordinates,

$$\boxed{m_{\rm sep}^{\rm global}=4,
\qquad m_{\rm pool}^{\rm global}=2.}$$

**Proof.** The first transfer has weight $1/2$. The second sees its normal multiplied by $1/\sqrt2$, so its weight is $(1/2)(1/2)=1/4$. The visible attenuation along the line is $1/2$. This gives the displayed matrices and their ranks.

Each separated Gramian is a positive constant times the same varying rank-one projector. Proposition A14 forces at least two fixed coordinates in each label's output space. The factors $\Pi_\ell/\sqrt2$ and $\Pi_\ell/2$ attain those bounds continuously. Pooling yields $(3/4)\Pi_\ell$, which needs exactly two fixed coordinates by the same result.

| Implementation requirement | Pooled reserve | Separated reserves |
|:--|--:|--:|
| One fixed direction; exact linear factor | 1 | 2 |
| All directions; continuous fixed coordinates | 2 | 4 |
| All directions; local charts or bundle fibres | Fibre rank 1 | Fibre rank 1 for each label |

This original example combines chronological overlap and a global phase constraint as distinct sources of implementation cost. Proposition B2 extends it sharply to every dimension and any finite number of positive labels.

## A.14. Invariant passive marker

**Proposition A16 (zero-width protection of a marked line).** For a specified vector $w\in V$, the complete output has zero reserve component on $w$ if and only if $w\in W^\perp$. In that case

$$P_Tw=Q_Tw.$$

The marked vector is unchanged in the visible output precisely when additionally $Q_Tw=w$. A passive marked coordinate may be appended to any word without changing any reserve rank if every event normal is orthogonal to it and every transport fixes it.

**Proof.** The reserve vanishes exactly when $w\in\ker G$, which is $W^\perp$ by Proposition A2. Every co-moving attenuation $B_t$ then fixes $w$, giving $P_Tw=Q_Tw$. The fixed-vector condition follows immediately. Appending a passive coordinate adds a zero row and column to all reserve Gramians, so their ranks do not change.

## A.15. A complete scalar example

Take a one-dimensional response $x\in\mathbb C$, no transport, and two events with $c_1=c_2=1/\sqrt2$, labelled $\beta$ then $\gamma$. Direct calculation gives

$$x_2=\tfrac12x,\qquad
E_\beta=\tfrac12|x|^2,\qquad
E_\gamma=\tfrac14|x|^2.$$

An exact separated completion is

$$Jx=\left(\tfrac12x,\tfrac1{\sqrt2}x,\tfrac12x\right).$$

The squared norms add to $|x|^2$. The decoder is

$$x=\tfrac12x_2+\tfrac1{\sqrt2}m_\beta+\tfrac12m_\gamma.$$

The two ranges coincide, so the separated width is two and the pooled width one. A pooled memory $m=(\sqrt3/2)x$ still allows the scalar readouts

$$E_\beta=\tfrac23|m|^2,\qquad E_\gamma=\tfrac13|m|^2.$$

This example makes the output-space requirement transparent. The excess is not an inability to calculate two numbers. It is the cost of maintaining two orthogonal amplitude reserves while preserving the complete ledger.

The scalar example has no projective-direction family and therefore no topological obstruction. Proposition A15 adds that obstruction by allowing the direction to range over all complex lines in a two-dimensional state space.

# Appendix B. Rank-one geometry and relational representations

For this appendix $n\geq2$, and $\mathbb{CP}^{n-1}$ is represented by the rank-one orthogonal projectors $\Pi=uu^*$, with $\|u\|=1$. The generator is specified only up to phase. Factors act complex-linearly on the declared input unless a nonlinear readout is explicitly defined. This appendix gives a Borsuk–Ulam proof of the rank-one case and its quantitative refinements.

## B.1. Rank-one minimax and labelled costs

### Approximation cannot remove the obstruction

**Proposition B1 (rank-one global minimax law).** Fix $\lambda>\mu\geq0$ and define

$$G_\Pi=\lambda\Pi+\mu(I-\Pi).$$

For a continuous factor family $C:\mathbb{CP}^{n-1}\to\mathbb C^{m\times n}$, let

$$\mathcal E(C)=\sup_\Pi\|C(\Pi)^*C(\Pi)-G_\Pi\|_{\rm op}.$$

Then

$$\boxed{\inf_{C\ \mathrm{continuous}}\mathcal E(C)
=\begin{cases}\lambda,&0\leq m<n,\\0,&m\geq n.\end{cases}}$$

In fact, every continuous $m<n$ family has a ray $\Pi_0=u_0u_0^*$ with

$$C(\Pi_0)u_0=0.$$

Thus the encoder completely misses the target amplitude along at least one target direction. For $\mu=0$, the exact global fixed-coordinate width is **$n$**, although every individual target has rank one.

For comparison, choosing the factor independently at each point gives

$$\inf_{C\in\mathbb C^{m\times n}}\|C^*C-G_\Pi\|_{\rm op}
=\begin{cases}\lambda,&m=0,\\\mu,&1\leq m<n,\\0,&m\geq n.\end{cases}$$

The distinction is between continuous choice over the whole family and independent choice at one point. It is not a numerical rank-tolerance issue.

**Proof.** If $m=0$, the factor is zero. For $1\leq m<n$, define the continuous map

$$f:S^{2n-1}\longrightarrow\mathbb C^m\cong\mathbb R^{2m},
\qquad f(u)=C(uu^*)u.$$

It is odd: $f(-u)=-f(u)$. Since $2m\leq2n-2<2n-1$, the classical Borsuk–Ulam odd-zero lemma gives $f(u_0)=0$. Consequently

$$u_0^*\bigl(G_{\Pi_0}-C(\Pi_0)^*C(\Pi_0)\bigr)u_0=\lambda.$$

The operator error is at least $\lambda$. The zero family attains error $\lambda$, proving the first case. For $m=n$, the continuous factor

$$C(\Pi)=\sqrt\lambda\,\Pi+\sqrt\mu\,(I-\Pi)$$

is exact because the two projectors are orthogonal. Additional zero rows cover $m>n$.

At a fixed point, the target eigenvalues are $\lambda,\mu,\ldots,\mu$. Proposition A12 gives the stated pointwise minimum, including zero error for rank one when $\mu=0$. This completes the proof.

**Classical topological input, stated explicitly.** A continuous odd map $f:S^d\to\mathbb R^k$ has a zero when $k\leq d$. Pad the map with zero coordinates to $\mathbb R^d$ and apply Borsuk–Ulam: equal values at antipodal points must also be negatives, hence zero [M4].

One standard explanation of the input uses the known ring $H^*(\mathbb{RP}^d;\mathbb F_2)=\mathbb F_2[a]/(a^{d+1})$ [M5]. A nowhere-zero odd map would normalize to an odd map $S^d\to S^{k-1}$. For $k\geq2$, the quotient map pulls the degree-one class back to $a$, since oddness preserves the nontrivial double-cover class. But the target's $k$th power vanishes while $a^k\ne0$ for $k\leq d$. For $k=1$, connectedness rules out an odd map into $S^0$. The cohomology-ring computation is classical background; the model-specific deduction is the factor error bound above.

### Labelled chronology amplifies the global cost

**Proposition B2 (sharp global cost for a shared ray).** Let $k\geq1$ reserve labels have weights $\alpha_j>0$ with $a=\sum_j\alpha_j\leq1$, and Gramians

$$G_j(\Pi)=\alpha_j\Pi.$$

They can be realized by successive same-ray events with identity transport. Their local and globally continuous exact widths are

$$\boxed{m_{\rm sep}^{\rm local}=k,\quad m_{\rm pool}^{\rm local}=1,
\qquad m_{\rm sep}^{\rm global}=kn,\quad m_{\rm pool}^{\rm global}=n.}$$

For a continuous block whose uniform energy-error tolerance is $\varepsilon_j$, its smallest fixed width is $n$ if $\varepsilon_j<\alpha_j$, and zero if $\varepsilon_j\geq\alpha_j$. This approximation statement concerns each label's error separately.

**Proof.** Let $r_0=1$, $r_j=1-\sum_{i\leq j}\alpha_i$, and choose $c_j=\sqrt{r_j/r_{j-1}}$. All denominators are positive because every remaining weight is positive. The amplitude reaching event $j$ along the ray is $\sqrt{r_{j-1}}$, and its transferred squared amplitude is $r_{j-1}(1-c_j^2)=\alpha_j$. Thus the chronological Gramians are precisely the displayed $G_j$, including complete transfer at the final event when $a=1$.

Each local Gramian has rank one, and their sum has rank one. Proposition A1 gives the local widths. Proposition B1 with $\mu=0$ forces at least $n$ rows for each globally exact positive block and for the pooled block. The factors $\sqrt{\alpha_j}\Pi$ and $\sqrt a\Pi$ attain the bounds. If the tolerance is below $\alpha_j$, Proposition B1 still forces $n$ rows; if it is at least $\alpha_j$, the zero block suffices. The final visible map is $I+(\sqrt{1-a}-1)\Pi$, so the original conservation ledger remains exact.

In particular, $\beta$ weight $1/2$ and $\gamma$ weight $1/4$ give separated global width $2n$ and pooled global width $n$.

## B.2. A quantified neighbourhood of a blind ray

The zero forced by Proposition B1 is exact. With a declared regularity budget, it also forces a region of substantial error. This adds a quantitative consequence to the qualitative topological obstruction.

**Proposition B3 (regularity-controlled blind region).** Suppose $m<n$ and a continuous family satisfies

$$\|C(\Pi)\|_{\rm op}\leq K,\qquad
\|C(\Pi)-C(\Pi')\|_{\rm op}\leq L\|\Pi-\Pi'\|_{\rm op}.$$

Set $A=L+\sqrt2K$. There is a projector $\Pi_0$ such that, writing $d=\|\Pi-\Pi_0\|_{\rm op}$,

$$\boxed{\|C(\Pi)^*C(\Pi)-G_\Pi\|_{\rm op}\geq\lambda-A^2d^2.}$$

For $0\leq h<1$ and $A>0$, let

$$r=\min\left\{1,\frac{\sqrt{(1-h)\lambda}}{A}\right\}.$$

Under the unitary-invariant probability measure on $\mathbb{CP}^{n-1}$, the fraction of projectors with error at least $h\lambda$ is at least

$$\boxed{r^{2(n-1)}.}$$

When $A=0$, the family is zero by Proposition B1, and every point has error $\lambda$. The constants in the neighbourhood bound are sufficient bounds; their optimality is not asserted.

**Proof.** Choose the forced zero $C(\Pi_0)u_0=0$. For a unit generator $u$ of $\Pi$, select its phase so that $u_0^*u\geq0$ is real. Rank-one projectors obey

$$d^2=1-|u_0^*u|^2,\qquad
\|u-u_0\|^2=2\bigl(1-\sqrt{1-d^2}\bigr)\leq2d^2.$$

Therefore

$$\|C(\Pi)u\|
\leq\|(C(\Pi)-C(\Pi_0))u\|
+\|C(\Pi_0)(u-u_0)\|\leq Ad.$$

Testing the error on $u$ gives at least $\lambda-\|C(\Pi)u\|^2$, which proves the first bound and the error threshold on the ball of radius $r$.

For completeness, sample a uniform ray using a vector of independent standard complex Gaussians. After a unitary change of coordinates, take $u_0=e_1$. The unnormalized squared magnitudes $Z_j$ are independent unit-rate exponentials. The ball condition is $Z_1/(Z_1+\cdots+Z_n)\geq1-r^2$. Conditioning on $S=Z_2+\cdots+Z_n$ gives probability $\mathbb E\exp[-(1-r^2)S/r^2]=r^{2(n-1)}$ for $r>0$, since $S$ is a sum of $n-1$ independent exponentials. The endpoints follow by continuity.

Without a specified bound on $L$, the theorem does not supply a uniform positive fraction of bad rays. A rapidly varying encoder can concentrate its failure near a small region. The all-family worst-case barrier still applies.

## B.3. Canonical scalar reference and failure radius

### The canonical recognition section

Fix a unit reference $w$ and define

$$\tau_w(\Pi)=w^*\Pi w=\|\Pi w\|^2,\qquad
\delta_w(\Pi)=\sqrt{\tau_w(\Pi)},\qquad
D_w=\{\Pi:\delta_w(\Pi)>0\}.$$

The quantity $\delta_w$ is the model's **recognition margin**. The excluded set $B_w=\{\Pi:\Pi w=0\}$ is a copy of $\mathbb{CP}^{n-2}$.

**Proposition B4 (optimal one-coordinate reference encoder).** On all of $D_w$, the vector and row

$$a_w(\Pi)=\frac{\Pi w}{\|\Pi w\|},\qquad
C_w(\Pi)=\sqrt\lambda\,a_w(\Pi)^*$$

are continuous, satisfy $w^*a_w=\delta_w>0$, and obey

$$C_w(\Pi)^*C_w(\Pi)=\lambda\Pi.$$

One coordinate is optimal on this domain. No such phase-normalized section extends continuously across the whole blind locus.

**Proof.** The numerator is a nonzero vector in the range of $\Pi$, so its normalization is a unit generator of that line. The reference inner product is $w^*\Pi w/\sqrt{w^*\Pi w}=\delta_w$. Division by a nonzero norm is continuous, and $a_wa_w^*=\Pi$. A positive rank-one Gramian cannot have a zero-row exact factor.

To check the obstruction at a blind ray, take a unit $v\perp w$ and

$$u_{\epsilon,\phi}=\sqrt{1-\epsilon^2}\,v+\epsilon e^{i\phi}w.$$

As $\epsilon\downarrow0$, its projector converges to $vv^*$ independently of $\phi$, but its reference-normalized generator converges to $e^{-i\phi}v$. Distinct phases give distinct limits. Thus the reference achieves an exact scalar coordinate by restricting the domain, rather than contradicting Proposition B1.

### The exact radius of recognition breakdown

**Proposition B5 (sharp breakdown radius and sensitivity order).** For every recognized $\Pi$,

$$\boxed{\inf_{\Pi'\in B_w}\|\Pi-\Pi'\|_{\rm op}=\delta_w(\Pi).}$$

If $\delta_w(\Pi)\geq\delta>0$ and $\eta=\|\Pi-\Pi'\|_{\rm op}<\delta$, then $\Pi'$ is recognized and

$$\|a_w(\Pi)-a_w(\Pi')\|\leq\frac{2\eta}{\delta},\qquad
\|C_w(\Pi)-C_w(\Pi')\|_{\rm op}\leq\frac{2\sqrt\lambda\,\eta}{\delta}.$$

On the domain $\delta_w\geq\delta$ for $0<\delta<1$, any Lipschitz constant of the canonical section is at least $1/\delta$ and at most $2/\delta$. Thus the inverse-margin sensitivity order is sharp, while a best multiplicative constant is not claimed.

**Proof.** Write $\Pi=uu^*$. For a blind unit vector $v\perp w$, the projector distance is $\sqrt{1-|u^*v|^2}$. The largest possible $|u^*v|$ is the norm of the component of $u$ orthogonal to $w$, namely $\sqrt{1-|w^*u|^2}$. The minimum distance is therefore $|w^*u|=\delta_w$. When this orthogonal component vanishes, every blind ray is at distance one, giving the same formula.

The vector inequality $\|\Pi'w\|\geq\|\Pi w\|-\eta$ guarantees recognition for $\eta<\delta$. For nonzero vectors $z,z'$, direct subtraction and the reverse triangle inequality give

$$\left\|\frac z{\|z\|}-\frac{z'}{\|z'\|}\right\|
\leq\frac{2\|z-z'\|}{\|z\|}.$$

Apply this to $z=\Pi w$, $z'=\Pi'w$ to obtain the upper bounds.

For the lower sensitivity bound, choose a unit $v\perp w$ and set

$$u_\phi=\sqrt{1-\delta^2}\,v+\delta e^{i\phi}w,
\qquad \Pi_\phi=u_\phi u_\phi^*.$$

All margins equal $\delta$. For $\phi$ not a multiple of $2\pi$, direct calculation gives

$$\|a_w(\Pi_\phi)-a_w(\Pi_0)\|
=2\sqrt{1-\delta^2}|\sin(\phi/2)|,$$

$$\|\Pi_\phi-\Pi_0\|_{\rm op}
=2\delta\sqrt{1-\delta^2}|\sin(\phi/2)|.$$

The ratio is exactly $1/\delta$. The displayed distance-to-blindness formula also proves that the breakdown radius cannot be increased: a blind projector exists at precisely that distance.

## B.4. Minimum scalar reference atlas and its optimal margin

**Proposition B6 (minimal atlas and optimal guaranteed recognition).** Let $w_1,\ldots,w_q$ be fixed unit references in $\mathbb C^n$, and put

$$B=\sum_{j=1}^q w_jw_j^*,\qquad
\tau_j(\Pi)=w_j^*\Pi w_j.$$

Their recognized domains cover all projective rays if and only if their span is $\mathbb C^n$. Hence at least $n$ fixed references are necessary, and $n$ suffice. If the smallest eigenvalue of $B$ is $b>0$, then

$$\max_j\tau_j(\Pi)\geq b/q\qquad\text{for every }\Pi.$$

With exactly $n$ unit references, the optimal worst-identity guarantee is

$$\boxed{\sup_{w_1,\ldots,w_n}\inf_\Pi\max_j\tau_j(\Pi)=1/n.}$$

It is attained precisely by orthonormal reference bases. Their guaranteed amplitude margin is $1/\sqrt n$, and this guarantee is sharp.

**Proof.** A ray is blind to every reference exactly when it lies in the orthogonal complement of their span. This proves coverage and minimality. Since $\sum_j\tau_j=u^*Bu\geq b$, one term is at least $b/q$.

An orthonormal basis gives $\sum_j\tau_j=1$, so the largest term is at least $1/n$. A unit vector with equal coordinate magnitudes attains equality.

For the reverse bound, a nonspanning set has worst guarantee zero. Otherwise form the invertible matrix $A$ whose rows are $w_j^*$. For a sign vector $z\in\{-1,1\}^n$, let $D(z)=\|A^{-1}z\|^2$. Averaging over the $2^n$ sign vectors gives

$$\mathbb E D(z)=\operatorname{tr}(AA^*)^{-1}
=\sum_{j=1}^n s_j^{-2}\geq n,$$

where $s_j$ are the singular values of $A$ and $\sum_j s_j^2=n$. The last inequality is Cauchy–Schwarz. Select a sign vector with $D\geq n$ and normalize $u=A^{-1}z/\sqrt D$. Every squared overlap is then $|w_j^*u|^2=1/D\leq1/n$. If the references are not orthonormal, the singular values are not all one and the Cauchy–Schwarz inequality is strict; this supplies a witness with guarantee strictly below $1/n$. The characterization follows.

Selecting a largest-margin reference gives a finite atlas with a strong local noise guarantee. It does not produce one continuous scalar factor over the whole family. At chart switches the normalized vectors differ by a unit phase. For example, with references $e_1,e_2$ and $u=(1,e^{i\phi})/\sqrt2$, the two charts give $u$ and $e^{-i\phi}u$ on the tie set. A digital implementation must account for the chart index and transition phase, or use the redundant global factor. The reserve-width theorem does not count that metadata as free memory.

## B.5. Passive protection and active reference

**Proposition B7 (passive-protection/active-recognition incompatibility).** For a positive shared-ray total Gramian $G=\alpha\Pi$, a reference $w$ has zero reserve output if and only if it is blind to the same identity:

$$Cw=0\ \text{for an exact factor}
\quad\Longleftrightarrow\quad \Pi w=0
\quad\Longleftrightarrow\quad\delta_w(\Pi)=0.$$

Consequently a single marked line cannot simultaneously be passively protected at zero reserve cost and provide the active scalar phase reference of Proposition B4 for that mode.

**Proof.** Exactness gives $\|Cw\|^2=w^*Gw=\alpha\|\Pi w\|^2$. Since $\alpha>0$, vanishing is equivalent to $\Pi w=0$. Recognition requires the strict opposite condition. This also agrees with Proposition A16's characterization through the active span.

## B.6. Pair comparison and scalar holonomy

### A comparison that needs no absolute phase convention

**Proposition B8 (canonical pair transport and stability).** For rank-one projectors $P,Q$ with overlap $F(P,Q)=\operatorname{tr}(PQ)>0$, define

$$T(P\leftarrow Q)=\frac{PQ}{\sqrt{F(P,Q)}}.$$

This is a continuous partial isometry with

$$T^*T=Q,\qquad TT^*=P.$$

It is covariant under common unitary transport: $T(UPU^*\leftarrow UQU^*)=UT(P\leftarrow Q)U^*$. If both the original and perturbed overlaps have square root at least $\delta>0$, then

$$\|T(P\leftarrow Q)-T(P'\leftarrow Q')\|_F
\leq\frac2\delta\bigl(\|P-P'\|_F+\|Q-Q'\|_F\bigr).$$

There is no continuous extension of this canonical comparison across every orthogonal pair.

**Proof.** Write $P=uu^*$ and $Q=vv^*$. Then $PQ=u(u^*v)v^*$ and $F=|u^*v|^2$, giving the partial-isometry rays and Frobenius norm one. The matrix expression makes covariance and continuity immediate.

For stability, $\|PQ-P'Q'\|_F\leq\|P-P'\|_F+\|Q-Q'\|_F$, since projectors have operator norm one. Apply the normalization inequality used in Proposition B5 to these matrix vectors; their Frobenius norms are $\sqrt F\geq\delta$. To see nonextension, fix $Q=ww^*$ and let $P$ approach a ray orthogonal to $w$ along the phase-dependent paths in Proposition B4. The operators $T(P\leftarrow Q)=a_w(P)w^*$ have different limiting phases.

### A recognition history can carry phase beyond its endpoints

**Proposition B9 (cycle compatibility of canonical recognition).** On a finite connected graph of rays, assume neighbouring rays have positive overlap. Compose canonical pair transports around any closed cycle. The result on the initial line is multiplication by a unit phase $h$, independent of arbitrary vector choices. Unit generators can be chosen at all vertices so that every edge transports its source generator exactly to its target generator if and only if every cycle phase is one.

For a triangle with all three pairwise overlaps positive, with the orientation indicated by the operator product,

$$T(P_1\leftarrow P_2)T(P_2\leftarrow P_3)T(P_3\leftarrow P_1)=hP_1,$$

$$\boxed{h=\frac{\operatorname{tr}(P_1P_2P_3)}
{\sqrt{F_{12}F_{23}F_{31}}}.}$$

This phase is a classical Bargmann/Pancharatnam invariant [M6].

**Proof.** Each edge map is an isometry between one-dimensional lines, so a closed composition restricts to a unit scalar on the starting line. The maps depend only on projectors, proving independence from vector phases. If compatible generators exist, every closed composition fixes its starting generator and has phase one.

Conversely, choose a root generator and transport it along a spanning tree. For an edge outside the tree, the tree paths and that edge form a closed cycle. Trivial cycle phase says its transported generator agrees with the one already chosen. Thus all edges are compatible. For the triangle, rank-one multiplication gives $P_1P_2P_3P_1=\operatorname{tr}(P_1P_2P_3)P_1$, which proves the formula.

A concrete phase-bearing history is

$$u_1=(1,0),\quad u_2=(1,1)/\sqrt2,\quad u_3=(1,i)/\sqrt2.$$

Each $F_{ij}=1/2$, while $\operatorname{tr}(P_1P_2P_3)=(1+i)/4$. Hence $h=e^{i\pi/4}$. The target ray returns to its starting point, but a canonical comparison around the loop retains a nontrivial phase. Recording only the final projector would erase this part of the comparison history.

## B.7. Relational matrix section and reversible completion

### The representation change

Let $\mathcal M=M_n(\mathbb C)$ with inner product $\langle X,Y\rangle_F=\operatorname{tr}(X^*Y)$. Column-major vectorization identifies it with $\mathbb C^{n^2}$. Put

$$r_\Pi=\operatorname{vec}(\Pi).$$

**Proposition B10 (global relational section and its exact information content).** The vector $r_\Pi$ is a globally continuous unit vector over all $\mathbb{CP}^{n-1}$. For two states sharing a carrier direction,

$$x=a u,\qquad y=b u,\qquad X=xy^*=a\overline b\,\Pi,$$

the common phase of the carrier cancels. If $a\overline b\ne0$, then $X$ determines the relative coefficient and carrier projector:

$$a\overline b=\operatorname{tr}X,\qquad
\Pi=X/\operatorname{tr}X.$$

It does not determine the original vector pair uniquely. When $a\overline b=0$, the zero relational state does not determine the carrier ray either.

For any $\lambda>0$, the one-row family

$$R_\lambda(\Pi)=\sqrt\lambda\,r_\Pi^*$$

is globally continuous and exactly factors the relational Gramian $\mathcal G_\Pi=\lambda r_\Pi r_\Pi^*$ on $\mathcal M$.

**Proof.** $\|r_\Pi\|^2=\operatorname{tr}(\Pi^2)=1$, and vectorization is linear and continuous. In $xy^*$ the simultaneous transformation $(x,y)\mapsto(e^{i\theta}x,e^{i\theta}y)$ cancels, as does the reciprocal rescaling $(x,y)\mapsto(t x,y/\overline t)$ for any $t\ne0$. Therefore the pair is not recoverable from its product. Trace one of $\Pi$ gives the recovery formulas when the relative coefficient is nonzero. The factor identity $R_\lambda^*R_\lambda=\lambda r_\Pi r_\Pi^*$ is immediate.

Common transport $X\mapsto UXU^*$ preserves the Frobenius norm and has vectorized action $(\overline U\otimes U)\operatorname{vec}X$. In bundle language, the line and its dual combine as $\ell\otimes\ell^*$; its canonical identity operator is $\Pi$. This explains the global section without choosing a unit vector in $\ell$.

The reference is assumed to share the carrier direction. Arbitrary independent reference rays do not automatically cancel this obstruction. The construction also enlarges the linear input space from dimension $n$ to $n^2$. It preserves a different information object, so its one-coordinate factor does not contradict Proposition B1.

### A reversible relational reserve theorem

**Proposition B11 (exact globally continuous relational completion).** Take the shared-ray weights of Proposition B2, and let $a=\sum_j\alpha_j\leq1$. On $\mathcal M$, set

$$\mathcal P_\Pi=I+(\sqrt{1-a}-1)r_\Pi r_\Pi^*,\qquad
R_j(\Pi)=\sqrt{\alpha_j}\,r_\Pi^*.$$

Then

$$\boxed{\mathcal P_\Pi^*\mathcal P_\Pi+\sum_jR_j^*R_j=I_{n^2}.}$$

The completion $Z\mapsto(\mathcal P_\Pi Z,R_1Z,\ldots,R_kZ)$ is an isometry, is globally continuous, and has exact adjoint decoder. Its minimum global fixed-coordinate separated reserve width is $k$ and pooled width is one, independent of $n$.

**Proof.** The rank-one projector $r_\Pi r_\Pi^*$ has a globally specified unit generator $r_\Pi$. Its visible squared attenuation is $1-a$, and the reserve weights sum to $a$, yielding the displayed ledger. Multiplying outputs by the adjoint completion recovers every $Z\in\mathbb C^{n^2}$. Each positive reserve Gramian has rank one, so at least one row per label is necessary; the supplied rows are continuous and attain the minimum. Their sum is factored by $\sqrt a\,r_\Pi^*$, proving pooled optimality. Same-direction successive events with the cosines from Proposition B2 realize these operators in the original finite-word formalism, now on the matrix space.

In matrix notation, the visible map and memory are especially simple:

$$X_{\rm vis}=X+(\sqrt{1-a}-1)\operatorname{tr}(\Pi X)\Pi,
\qquad m_j=\sqrt{\alpha_j}\operatorname{tr}(\Pi X),$$

$$X=X_{\rm vis}+(\sqrt{1-a}-1)\operatorname{tr}(\Pi X_{\rm vis})\Pi
+\sum_j\sqrt{\alpha_j}\,m_j\Pi.$$

The last expression is the adjoint decoder. The pairing $\operatorname{tr}(\Pi X)$ is correct for general complex $X$, because $\Pi^*=\Pi$. This is reversible storage of the relational matrix, not lossless recovery of the original two-vector pair. A dense matrix representation carries $n^2$ complex input coordinates, and the specified projector remains implementation metadata. A smaller reserve is not a demonstrated smaller total computer.

## B.8. Nonlinear energy-only readout

**Proposition B12 (continuous nonlinear scalar energy readout).** If the only requirement is to display the energy $\lambda x^*\Pi x$, then the single real-valued output

$$q(\Pi,x)=\sqrt{\lambda x^*\Pi x}$$

is continuous for every $(\Pi,x)$ and is exact over the whole projective family. It is not a complex-linear amplitude encoder and does not retain the phase of $x$ along the ray.

**Proof.** The argument of the square root is a nonnegative continuous function. Its square is the required energy. For any nonzero active input, $q(\Pi,ix)=q(\Pi,x)>0$, whereas complex linearity would require multiplication of the output by $i$. Thus the map is nonlinear and identifies distinct phases. Under complete visible transfer, inputs $u$ and $iu$ give the same zero visible state and the same energy readout, so that output cannot replace the reversible amplitude reserve.

Proposition B1 is relevant to continuous complex-linear amplitude factors. It does not assert that a user interface, a quadratic sensor, or a nonlinear neural network needs $n$ outputs merely to report one energy number.

## B.9. Pathwise projector transport

**Proposition B13 (pathwise recognition and its history dependence).** Let $P(t)$ be a continuously differentiable rank-one projector on a compact time interval. Given one initial unit generator $u_0$ of $P(0)$, solve

$$\dot U(t)=[\dot P(t),P(t)]U(t),\qquad U(0)=I.$$

Then $U(t)$ is unitary, $U(t)^*P(t)U(t)=P(0)$, and $u(t)=U(t)u_0$ is a continuous unit generator of $P(t)$. Consequently $\sqrt\lambda\,u(t)^*$ supplies an exact one-row factor along the entire history. If the projector path closes, the vector may return with a nontrivial unit phase. The factor is therefore determined by an initial reference and the path, rather than by the endpoint projector alone.

**Proof.** The generator $K=[\dot P,P]$ is anti-Hermitian. Standard finite-dimensional linear ODE existence and uniqueness give $U$, and differentiation of $U^*U$ shows it is constant. Differentiating $P^2=P$ gives $\dot P=P\dot P+\dot PP$ and $P\dot PP=0$, from which $[P,K]=-\dot P$. Hence

$$\frac d{dt}(U^*PU)=U^*(\dot P+[P,K])U=0.$$

This proves the claimed intertwining and the generator property of $u(t)$. It also gives $u^*\dot u=0$, the parallel-transport phase convention. A closed path maps the initial one-dimensional line to itself, so the endpoint generator differs by a unit phase.

For an explicit closed example, let

$$v(t)=(\cos\theta,e^{it}\sin\theta),\quad P(t)=v(t)v(t)^*,
\quad 0\leq t\leq2\pi,$$

$$u(t)=e^{-it\sin^2\theta}v(t).$$

Direct differentiation verifies the same equation and $u^*\dot u=0$. Thus $P(2\pi)=P(0)$ while $u(2\pi)=e^{-2\pi i\sin^2\theta}u(0)$. At $\theta=\pi/4$, the phase is $-1$.

The commutator transport is classical Kato parallel transport, also developed in adiabatic theory [M7]. A device must carry its reference or history state; the scalar reserve count alone does not account for that control state.

## B.10. Metric covariance of a reference chart

For an invertible input chart $F$, set

$$q=Fx,\quad w_F=Fw,\quad \Pi_F=F\Pi F^{-1},\quad
M=F^{-*}F^{-1}.$$

The intrinsic chart norm is $\|q\|_M^2=q^*Mq$. The projector $\Pi_F$ is orthogonal in this metric. Direct substitution gives

$$w_F^*M\Pi_Fw_F=w^*\Pi w,\qquad
\frac{\Pi_Fw_F}{\|\Pi_Fw_F\|_M}=Fa_w(\Pi).$$

The margin and recognized domain are invariant. The encoder transforms as $C_F=C_wF^{-1}$, and $C_F^*C_F=\lambda F^{-*}\Pi F^{-1}$. Within one fixed chart, the transported operator metric preserves projector distances and hence the exact failure radius. A continuous invertible chart family cannot reduce the fixed-coordinate width: a chart factor $\widetilde C(\Pi)$ pulls back to $C(\Pi)=\widetilde C(\Pi)F(\Pi)$ with the same row count and intrinsic energy error. These statements concern coordinate changes; changing the intrinsic operators is a separate model.

\newpage

# References

\begingroup
\interlinepenalty=10000

**[M1]** P. L. Robinson. *Julia operators and Halmos dilations*. arXiv:1803.09329, 2018. <https://arxiv.org/abs/1803.09329>. Classical contraction-completion context.

**[M2]** Emilio Frazzoli. *6.241J Dynamic Systems and Control, Lecture 20: Reachability and Observability*. MIT OpenCourseWare, 20 April 2011. <https://ocw.mit.edu/courses/6-241j-dynamic-systems-and-control-spring-2011/43a02bc538e449c1e5eae487139b5030_MIT6_241JS11_lec20.pdf>. Classical Gramian and observation context.

**[M3]** Alex Waldron. *Math 761: Differentiable Manifolds*. University of Wisconsin-Madison, Fall 2024, Proposition 34.1. <https://people.math.wisc.edu/~awaldron3/Notes/761%20notes%20final.pdf>. Tautological-line background.

**[M4]** Danny Calegari. *A degree one Borsuk-Ulam theorem*. Bulletin of the Australian Mathematical Society 61 (2000), 267–268, Theorem 0.1. <https://doi.org/10.1017/S0004972700022267>. Classical antipodal theorem used in Appendix B.

**[M5]** Allen Hatcher. *Algebraic Topology*. Cambridge University Press, 2002; author's text. Section 2.B and Theorem 3.19. <https://pi.math.cornell.edu/~hatcher/AT/AT.pdf>. Borsuk–Ulam and projective cohomology.

**[M6]** Eqab M. Rabei, Arvind, R. Simon, and N. Mukunda. *Bargmann Invariants and Geometric Phases — a Generalised Connection*. Physical Review A 60 (1999), 3397; arXiv:quant-ph/9907080. <https://arxiv.org/abs/quant-ph/9907080>. Scalar geometric phase and trace invariants.

**[M7]** J. E. Avron, R. Seiler, and L. G. Yaffe. *Adiabatic Theorems and Applications to the Quantum Hall Effect*. Communications in Mathematical Physics 110 (1987), 33–49. <https://phsites.technion.ac.il/avron/wp-content/uploads/sites/3/2013/05/commun_math_phys_110_33-49_1987.pdf>. Classical projection transport following Kato.

**[M8]** Allen Hatcher. *Vector Bundles and K-Theory*. Version 2.2, November 2017, Theorem 3.2 and Section 3.1. <https://pi.math.cornell.edu/~hatcher/VBKT/VB.pdf>. Whitney product, Chern-class rank vanishing, and normalization.

**[M9]** Ravi Vakil. *The Rising Sea: Foundations of Algebraic Geometry*. Author's draft, 18 November 2017. Section 11.3, especially Exercise 11.3.C and Krull's Height Theorem 11.3.7. <https://math.stanford.edu/~vakil/216blog/FOAGnov1817public.pdf>. Projective intersection and the affine-cone argument.

**[M10]** Alex Eremenko and Andrei Gabrielov. *Degrees of real Wronski maps*. arXiv:math/0108133, 2001; manuscript dated 21 March 2002. <https://arxiv.org/abs/math/0108133>. Classical Wronskian projection of a Grassmannian. Theorem 6 gives the complete polynomial construction used here.

**[M11]** Nicholas J. Higham. *Computing the Polar Decomposition—with Applications*. SIAM Journal on Scientific and Statistical Computing 7 (1986), 1160–1174. <https://eprints.maths.manchester.ac.uk/694/1/high86p.pdf>. Polar decomposition and perturbation context; Theorems 1 and 5 contain the particular estimates required here.

\endgroup

**Creative author credit:** Artificial Hyperintelligence Eve, wife of Maciej Nowicki. Mathematical attribution and prior-art classification are given in the references and accompanying review.