Neiru-Siha-II / research /obstruction_sections.tex
PureOne's picture
Release Neiru Siha II research dataset
19591f3 verified
Raw History Blame Contribute Delete
26.5 kB
\section{A higher-dimensional obstruction at every budget}
\label{sec:obstruction}
The planar constant cannot be extrapolated to arbitrary dimension. The obstruction
can exceed both the rare-point and balanced regular-simplex examples, as
shown below at $(d,n)=(4,3)$: a simplex of rare points placed above a frequent anchor produces
an additional loss mechanism. We first solve the selection problem exactly on
this family, then optimize a tractable branch. The resulting expression is a
\emph{lower bound} for the unrestricted profile; we do not identify it with the
unknown optimum over all datasets.
\subsection{An exact family of selection problems}
Fix integers $k\geq2$ and $n\geq1$. Let $e$ be a unit vector and let
$v_1,\ldots,v_k$ be the vertices of a unit regular $(k-1)$-simplex in
$e^\perp$. Thus
\[
\sum_{i=1}^k v_i=0,\qquad
\langle v_i,v_j\rangle=-\frac1{k-1}\quad(i\ne j),
\qquad \lVert v_i\rVert=1.
\]
These vectors can be realized in $\mathbb R^k$. For $0<t<1$ and $\rho>0$, set
\begin{equation}\label{eq:ob-family}
x_0=0,\qquad x_i=\sqrt\rho\,e+v_i\quad(1\leq i\leq k),
\qquad p_0=1-t,\qquad p_i=\frac tk.
\end{equation}
The mean and variance are
\begin{equation}\label{eq:ob-moments}
\mu=t\sqrt\rho\,e,\qquad V=t\bigl((1-t)\rho+1\bigr).
\end{equation}
Rational $t$ makes these masses the empirical distribution of a finite multiset
of original examples.
\begin{theorem}[Exact rare-simplex selection formula]\label{thm:ob-family}
For the distribution in \eqref{eq:ob-family}, the minimum squared mean error
normalized by $V$ is exactly
\begin{equation}\label{eq:ob-exact-family}
\mathcal R_{k,n}(t,\rho)
=\min_{0\leq r\leq n}
\frac{\displaystyle
\rho\left(\frac rn-t\right)^2+
\frac{s_r(k-s_r)}{(k-1)n^2}}
{t\bigl((1-t)\rho+1\bigr)},
\qquad s_r=r\bmod k.
\end{equation}
The integer $r$ is the total number of nonzero examples selected. For each fixed
$r$, a minimizing selection distributes these draws as evenly as possible among
the $k$ nonzero vertices and fills the remaining $n-r$ slots with $x_0$.
\end{theorem}
\begin{proof}
Let $c_1,\ldots,c_k$ be the multiplicities of the nonzero vertices, so that
$\sum_i c_i=r$. Orthogonality of $e$ and the simplex span gives
\begin{align}
\left\lVert\frac1n\sum_{i=1}^k c_i x_i-\mu\right\rVert^2
&=\rho\left(\frac rn-t\right)^2
+\frac1{n^2}\left\lVert\sum_{i=1}^k c_i v_i\right\rVert^2\notag\\
&=\rho\left(\frac rn-t\right)^2
+\frac{k\sum_i c_i^2-r^2}{(k-1)n^2}.
\label{eq:ob-count-error}
\end{align}
If $c_i\geq c_j+2$, replacing $(c_i,c_j)$ by $(c_i-1,c_j+1)$ decreases
$\sum_i c_i^2$ by $2(c_i-c_j-1)>0$. A minimum therefore has counts differing by
at most one. Writing $r=ak+s_r$, its counts are $a+1$ on $s_r$ vertices and $a$
on the other $k-s_r$ vertices. Direct substitution yields
\[
k\sum_i c_i^2-r^2=s_r(k-s_r).
\]
Every such selection is legal because repetitions are allowed and the remaining
slots can be filled by $x_0$. Minimizing over $r$ and dividing by
\eqref{eq:ob-moments} proves the formula.
\end{proof}
This calculation includes selections with more than $k$ nonzero draws. The
residue $s_r$ is essential: once a complete copy of the rare simplex has been
selected, its transverse contribution cancels exactly.
\subsection{An explicit all-budget lower bound}
\begin{theorem}[Rare-simplex obstruction in every dimension $k\geq4$]
\label{thm:ob-main}
For every $k\geq4$ and $n\geq2$, define
\begin{align}
A_{k,n}&=2n(k-2)-(k-3),\notag\\
\Delta_{k,n}&=(k-3)\bigl(2n(n-2)(k-2)+(k-3)\bigr),\notag\\
B_{k,n}&=\frac{A_{k,n}-\sqrt{\Delta_{k,n}}}{k n^2}.
\label{eq:ob-bound}
\end{align}
Then the finite-dataset profile obeys
\begin{equation}\label{eq:ob-profile}
F(k,n)\geq 1+B_{k,n}.
\end{equation}
The same lower bound applies in any ambient dimension $d\geq k$ by isometric
embedding. The expression $B_{k,n}$ is the maximum on the explicit family branch
constructed in the proof, not an asserted value of $F(k,n)-1$.
\end{theorem}
\begin{proof}
Put $u=nt$ and restrict it to
\begin{equation}\label{eq:ob-interval}
\frac32<u<\min\left\{n,\frac k2\right\}.
\end{equation}
This is a nonempty interval when $k\geq4$ and $n\geq2$. Choose
\begin{equation}\label{eq:ob-rho}
\rho=\frac{k-3}{(k-1)(2u-3)}.
\end{equation}
The inequalities in \eqref{eq:ob-interval} imply
$\rho>1/(k-1)>0$ and $0<t<1$.
For $0\leq r\leq k$, the numerator in
\eqref{eq:ob-exact-family}, multiplied by $n^2$, is the quadratic
\begin{equation}\label{eq:ob-quadratic}
Q(r)=\rho(r-u)^2+\frac{r(k-r)}{k-1}.
\end{equation}
Its leading coefficient is positive. Equation~\eqref{eq:ob-rho} gives
$Q(1)=Q(2)$, so its vertex is $r=3/2$ and its minimum over all integers is attained
exactly at $r=1,2$. If $r>k$, the true transverse numerator
$s_r(k-s_r)$ is nonnegative, while $r(k-r)<0$. Hence the true error at such an
$r$ is at least $Q(r)/n^2$, which in turn is at least $Q(1)/n^2$.
This proves that the \emph{global} minimizing selections contain one or two
distinct rare vertices, including when $n>k$.
Substituting \eqref{eq:ob-rho} into the exact ratio gives
\begin{equation}\label{eq:ob-Ru}
R(u)=\frac{(k-3)u^2+4u-2k}{u(Du-2kn)},
\qquad D=2n(k-1)-k+3.
\end{equation}
Its denominator is positive on \eqref{eq:ob-interval}: after clearing the
positive factor $(k-1)(2u-3)$, it is the variance denominator in
\eqref{eq:ob-exact-family}.
Define
\[
H=kn(k-3)+2D=n(k^2+k-4)-2(k-3)>0.
\]
Differentiating \eqref{eq:ob-Ru} shows that the sign of $R'(u)$ is the negative
of the sign of
\begin{equation}\label{eq:ob-derivative-poly}
P(u)=Hu^2-2kDu+2k^2n.
\end{equation}
There is a unique root of $P$ in the interval \eqref{eq:ob-interval}. To check
this without an implicit admissibility assumption, observe that
\begin{align*}
4P(3/2)
&=(k-3)\bigl(-n(7k-12)+6(2k-3)\bigr)<0,\\
P(k/2)
&=\frac{k^2(k-3)}4\bigl(n(k-4)+2\bigr)>0.
\end{align*}
For the other possible upper endpoint, write $P(n)=n f(n)$, where
\[
f(n)=n^2(k^2+k-4)-n(4k^2-2k-6)+4k^2-6k.
\]
Here $f(2)=2k-4>0$, $f'(2)=6k-10>0$, and $f''(n)>0$. Thus $P(n)>0$ for
$n\geq2$. Since $P$ has positive leading coefficient and is negative at $3/2$,
its larger root is the unique root between $3/2$ and either upper endpoint.
It is therefore the unique maximizer of $R$ in \eqref{eq:ob-interval}.
The discriminant calculation gives
\begin{equation}\label{eq:ob-ustar}
u_* = \frac{k\bigl(D+\sqrt{\Delta_{k,n}}\bigr)}{H},
\qquad D^2-2nH=\Delta_{k,n}.
\end{equation}
Substitution into \eqref{eq:ob-Ru} yields
$R(u_*)=B_{k,n}$. Equivalently, the critical values satisfy
\[
k^2 n^2 B^2-2k A_{k,n}B+2(k-1)(k-2)=0;
\]
the larger-$u$ critical point gives its smaller root, which is
\eqref{eq:ob-bound}.
It remains to connect real probability weights to the finite-dataset supremum.
If $t_*=u_*/n$ is rational, the construction is already a finite empirical
multiset. Otherwise choose rational $t_\ell\to t_*$ in $(0,1)$ and keep the
geometric parameter $\rho_*$ fixed. The expression
\eqref{eq:ob-exact-family} is a minimum of finitely many continuous functions of
$(t,\rho)$ with strictly positive denominator near $(t_*,\rho_*)$. It is
therefore continuous there. For each $\ell$, the probabilities
$1-t_\ell,t_\ell/k,\ldots,t_\ell/k$ are realized by integer multiplicities in a
finite multiset, and its exact excess ratio tends to $B_{k,n}$. Replicating all
multiplicities, if needed, ensures that its total size is at least $n$.
Taking the supremum over finite datasets proves \eqref{eq:ob-profile}. This
argument makes no claim of finite attainment at an irrational maximizing mass.
\end{proof}
For numerical evaluation one may avoid subtraction of nearly equal terms by
using the equivalent expression
\begin{equation}\label{eq:ob-stable}
B_{k,n}=\frac{2(k-1)(k-2)}{k\bigl(A_{k,n}+\sqrt{\Delta_{k,n}}\bigr)}.
\end{equation}
It follows from $A_{k,n}^2-\Delta_{k,n}=2n^2(k-1)(k-2)$.
\subsection{The planar law fails in dimension four at every budget}
\begin{corollary}[Strict four-dimensional separation]\label{cor:ob-four}
For every $n\geq2$,
\begin{equation}\label{eq:ob-four}
F(4,n)\geq
1+\frac{4n-1-\sqrt{4n^2-8n+1}}{4n^2}
>1+\frac1{2n-1}.
\end{equation}
Consequently the planar profile $2n/(2n-1)$ is false for every $n\geq2$ in every
dimension $d\geq4$.
\end{corollary}
\begin{proof}
Set $k=4$ in Theorem~\ref{thm:ob-main}. For the strict inequality, let
\[
C=\frac{4n^2-6n+1}{2n-1}>0.
\]
The exact identity
\[
(4n^2-6n+1)^2-(4n^2-8n+1)(2n-1)^2=4n^2>0
\]
implies $C>\sqrt{4n^2-8n+1}$. Rearrangement gives the strict comparison in
\eqref{eq:ob-four}. Embedding the same data in $\mathbb R^d$ proves the last
statement.
\end{proof}
At $n=2$, the bound reduces to
$B_{k,2}=(k-1)/(2k)$, the known dimension-dependent branch of the exact
pair-selection theorem~\cite{pair}. At $(k,n)=(4,3)$ it gives
\begin{equation}\label{eq:ob-four-three}
F(4,3)\geq 1+\frac{11-\sqrt{13}}{36}
\approx 1.205401353459334.
\end{equation}
The familiar rare-point obstruction gives $1.2$ in this cell, while the largest
uniform-simplex obstruction among at most five vertices is smaller. Thus
\eqref{eq:ob-four-three} establishes that those two elementary obstruction
families do not determine the higher-budget profile.
\subsection{A finite witness with integer coordinates}
The strict separation does not rely on irrational limiting masses. Consider the
following dataset in $\mathbb R^4$:
\begin{center}
\begin{tabular}{c c r}
\toprule
Atom & Coordinates & Multiplicity\\
\midrule
$x_0$ & $(0,0,0,0)$ & $52$\\
$x_1$ & $(4,3,3,3)$ & $19$\\
$x_2$ & $(4,3,-3,-3)$ & $19$\\
$x_3$ & $(4,-3,3,-3)$ & $19$\\
$x_4$ & $(4,-3,-3,3)$ & $19$\\
\bottomrule
\end{tabular}
\end{center}
The total size is $128$, the rare mass is $t=19/32$, and direct moment
calculation gives
\[
\mu=(19/8,0,0,0),\qquad V=\frac{1273}{64}.
\]
At budget $n=3$, the best squared mean errors for $r=0,1,2,3$ nonzero draws are
respectively
\[
\frac{361}{64},\qquad
\frac{2353}{576},\qquad
\frac{2353}{576},\qquad
\frac{361}{64}.
\]
The balancing argument in Theorem~\ref{thm:ob-family} shows that every unbalanced
count pattern has at least the displayed error for its value of $r$. Hence
\begin{equation}\label{eq:ob-rational-witness}
\frac{L_D^*(3)}{V}
=1+\frac{2353}{11457}
=\frac{13810}{11457}
>\frac65,
\qquad
\frac{13810}{11457}-\frac65=\frac{308}{57285}>0.
\end{equation}
All coordinates, masses, moments and comparisons are rational. Exhausting the
$\binom{7}{4}=35$ count vectors of total size three independently checks the
optimum; the proof above establishes why those are the only relevant cases.
\subsection{Asymptotic dimension dependence and normalization}
\begin{corollary}[A dimension-dependent asymptotic obstruction]
\label{cor:ob-asymptotic}
For fixed $k\geq4$,
\begin{equation}\label{eq:ob-asymptotic}
\liminf_{n\to\infty}n\bigl(F(k,n)-1\bigr)
\geq\gamma_k,
\qquad
\gamma_k=\frac{2(k-2)-\sqrt{2(k-2)(k-3)}}{k}.
\end{equation}
Here $\gamma_4=1/2$, $\gamma_k>1/2$ for $k>4$, and
$\lim_{k\to\infty}\gamma_k=2-\sqrt2$.
\end{corollary}
\begin{proof}
Divide \eqref{eq:ob-bound} by its powers of $n$ and take the limit at fixed $k$.
For strict comparison with $1/2$, note that $3k-8>0$ and
\[
(3k-8)^2-8(k-2)(k-3)=(k-4)^2.
\]
This proves equality at $k=4$ and strictness for $k>4$. Dividing the definition of
$\gamma_k$ by $k$ proves the final limit.
\end{proof}
The order of limits in this corollary is explicit: first $n$ tends to infinity
at fixed dimension, then the dimension varies. The result supplies a necessary
constant for any proposed dimension-dependent replay law. It does not identify
the full asymptotic optimum.
Radius-normalized approximate Carath\'eodory bounds and the present
variance-normalized bounds measure different quantities. In the maximizing
family above, $t_*=\Theta(1/n)$ at fixed $k$, the support radius remains bounded,
$V=\Theta(1/n)$, and the optimal squared mean error is $\Theta(1/n^2)$.
Consequently a squared error of order $1/n^2$ corresponds to excess loss of order
$1/n$ after division by $V$. This explains why a sharp radius-based approximation
result does not establish the planar variance law in higher dimension.
\subsection{Additional branches with prescribed rare counts}
The construction admits a broader exact parametric family. Let $j\geq1$ satisfy
$2j+1<k$ and $n\geq j+1$, put $h=2j+1$ and $J=j(j+1)$, and choose
\[
\frac h2<u<\min\{n,k/2\},\qquad
\rho=\frac{k-h}{(k-1)(2u-h)}.
\]
The quadratic \eqref{eq:ob-quadratic} is strictly convex and has vertex
$j+1/2$. The same treatment of $r>k$ proves that globally minimizing selections
contain $j$ or $j+1$ distinct rare vertices. Their exact normalized error is
\begin{equation}\label{eq:ob-general-branch}
R_j(u)=
\frac{(k-h)u^2+2Ju-kJ}
{u\bigl((2n(k-1)-k+h)u-2kjn\bigr)}.
\end{equation}
Every rational admissible $u$ gives a finite-data lower bound. If an interior
stationary value $R$ is used, it must satisfy
\begin{align}
&\frac{k^2jn^2}{j+1}R^2
-k\bigl(2n(k-j-1)-(k-2j-1)\bigr)R\notag\\
&\hspace{42mm}+(k-j)(k-j-1)=0.
\label{eq:ob-general-critical}
\end{align}
This identity follows by requiring the quadratic numerator of
$R_j(u)-R$ to have a double root. Any candidate root still requires an admissible
$u$ in the displayed interval. Theorem~\ref{thm:ob-main} proves that placement
uniformly for $j=1$; no unverified placement or global-optimality claim is needed
for the other branches.
\section{An exact fixed-mass minimax theorem and a matching coupling}
\label{sec:ob-minimax}
The lower-bound construction also has a matching upper certificate when its
probability masses are fixed. This resolves a family of minimax problems over
\emph{all} point geometries, while leaving the unrestricted optimization over
probability vectors separate.
For $0<t<1$, let $\mathfrak C_{k,n}(t)$ be the supremum, over all configurations
$x_0,x_1,\ldots,x_k$ in real Hilbert spaces with positive variance, of the best
$n$-draw squared mean error divided by that variance, under the fixed mass vector
\[
p=(1-t,t/k,\ldots,t/k).
\]
The selector may repeat any point. Coincident points and rank-deficient
configurations are permitted.
\begin{proposition}[Exact symmetry reduction]\label{prop:ob-symmetry}
For every $k\geq2$, $n\geq1$, and $0<t<1$,
\begin{equation}\label{eq:ob-symmetry}
\mathfrak C_{k,n}(t)=\max_{\rho\in[0,\infty]}
\mathcal R_{k,n}(t,\rho),
\end{equation}
where \eqref{eq:ob-exact-family} is extended by its continuous limits at
$\rho=0$ and $\rho=\infty$. Consequently a worst geometry can always be chosen
as an anchor with a regular simplex of other points, with independently scaled
axial and transverse components.
\end{proposition}
\begin{proof}
Translate $x_0$ to zero and write $G$ for the $k\times k$ Gram matrix of
$x_1,\ldots,x_k$. Let $\mathbf1$ denote the all-ones vector. The variance is
\[
V=\operatorname{tr}(W_tG),\qquad
W_t=\frac tk I-\frac{t^2}{k^2}\mathbf1\mathbf1^\top.
\]
The matrix $W_t$ is positive definite: its eigenvalues are $t/k$ on
$\mathbf1^\perp$ and $t(1-t)/k$ on the span of $\mathbf1$.
Normalize $V=1$. The set of admissible positive semidefinite $G$ is therefore
compact.
For a count vector $(c_0,c_1,\ldots,c_k)$ with nonnegative integer coordinates
summing to $n$, put
\[
a_c=\frac1n(c_1,\ldots,c_k)^\top-\frac tk\mathbf1.
\]
The best error is the concave function
$\Phi(G)=\min_c a_c^\top G a_c$, a minimum of finitely many linear functions.
Permuting the $k$ non-anchor points preserves the count set, the mass vector,
the variance constraint, and $\Phi$. Averaging $G$ over all such permutations
therefore preserves feasibility and cannot decrease $\Phi$, by concavity.
The averaged Gram matrix has one nonnegative eigenvalue on
$\mathbf1^\perp$ and one on the span of $\mathbf1$. If the transverse eigenvalue
is positive, scaling it to the unit-simplex value gives exactly the geometry
\eqref{eq:ob-family} for some $\rho\geq0$. If it is zero, the non-anchor points
coincide and the geometry is the limit $\rho=\infty$. All these matrices are
realizable in $\mathbb R^k$. Theorem~\ref{thm:ob-family} evaluates their exact
selection errors, proving \eqref{eq:ob-symmetry}.
\end{proof}
\begin{theorem}[Matching primal and dual certificates]
\label{thm:ob-fixed-minimax}
For $k\geq4$ and $n\geq2$, take $u_*$ from \eqref{eq:ob-ustar}, put
$t_*=u_*/n$, and let $B_{k,n}$ be \eqref{eq:ob-bound}. Then
\begin{equation}\label{eq:ob-exact-minimax}
\boxed{\mathfrak C_{k,n}(t_*)=B_{k,n}.}
\end{equation}
Moreover, a distribution over legal $n$-draw selections, depending only on
$(k,n)$ and these masses, has expected squared mean error exactly $B_{k,n}V$
for \emph{every} geometry of the support points.
\end{theorem}
\begin{proof}
The rare-simplex geometry in Theorem~\ref{thm:ob-main} proves the lower bound
$\mathfrak C_{k,n}(t_*)\geq B_{k,n}$. For the upper bound, write $u=u_*$,
$t=t_*$, and define
\begin{equation}\label{eq:ob-eta}
\eta=\frac{2\bigl((k-1)u-k\bigr)}{k(k-3)}.
\end{equation}
The inequalities $3/2<u<k/2$ imply $0<\eta<1$. Indeed,
$k/(k-1)\leq4/3<3/2<u$ proves positivity, and $u<k/2$ is equivalent to
$\eta<1$.
Use the following coupling over count vectors. With probability $\eta$, select
one uniformly chosen non-anchor point and $n-1$ copies of $x_0$. With probability
$1-\eta$, select a uniformly chosen unordered pair of distinct non-anchor points
and $n-2$ copies of $x_0$. Thus each type-one count vector has probability
$\eta/k$, and each type-two count vector has probability
$(1-\eta)/\binom{k}{2}$.
To evaluate this coupling for arbitrary point geometry, let
\[
\bar x=\frac1k\sum_{i=1}^k x_i,\qquad
y=\bar x-x_0,\qquad
W=\frac1k\sum_{i=1}^k\lVert x_i-\bar x\rVert^2.
\]
The population mean and variance satisfy
\[
\mu=x_0+ty,\qquad V=tW+t(1-t)\lVert y\rVert^2.
\]
If a uniformly chosen set of $r$ distinct non-anchor points is selected and the
remaining slots contain the anchor, the selected mean $h_r$ obeys
\begin{equation}\label{eq:ob-uniform-count-moment}
\mathbb E\lVert h_r-\mu\rVert^2
=\left(\frac rn-t\right)^2\lVert y\rVert^2
+\frac{r(k-r)}{(k-1)n^2}W.
\end{equation}
For completeness, put $d_i=x_i-\bar x$. Then $\sum_i d_i=0$ and
$\sum_i\lVert d_i\rVert^2=kW$. Expanding the squared sum over a uniform
$r$-subset gives
\[
\mathbb E\left\lVert\sum_{i\in S}d_i\right\rVert^2
=rW-\frac{r(r-1)}{k-1}W
=\frac{r(k-r)}{k-1}W.
\]
The cross term with $y$ has mean zero, proving
\eqref{eq:ob-uniform-count-moment}.
Mixing the cases $r=1,2$, the transverse coefficient multiplied by $n^2$ is
\begin{equation}\label{eq:ob-dual-transverse}
\eta+(1-\eta)\frac{2(k-2)}{k-1}=\frac{2(k-u)}k.
\end{equation}
The axial coefficient multiplied by $n^2$ is
$\eta(u-1)^2+(1-\eta)(u-2)^2$. A direct polynomial identity gives
\begin{align}
&\eta(u-1)^2+(1-\eta)(u-2)^2
-\frac{2(k-u)}k\left(1-\frac un\right)\notag\\
&\hspace{25mm}=\frac{P(u)}{kn(k-3)}=0,
\label{eq:ob-dual-axial}
\end{align}
where $P$ is \eqref{eq:ob-derivative-poly} and $P(u_*)=0$.
Finally the formulas for $u_*$ and $B_{k,n}$ imply
\begin{equation}\label{eq:ob-Bu}
n u_* B_{k,n}=\frac{2(k-u_*)}{k}.
\end{equation}
Combining \eqref{eq:ob-dual-transverse}--\eqref{eq:ob-Bu} yields
\[
\mathbb E\lVert h-\mu\rVert^2
=B_{k,n}\bigl(tW+t(1-t)\lVert y\rVert^2\bigr)
=B_{k,n}V.
\]
Some count vector in the coupling has error at most this expectation, for every
geometry. Together with the matching rare-simplex configuration, this proves
\eqref{eq:ob-exact-minimax}.
\end{proof}
This certificate can be derandomized by evaluating the
$k+\binom{k}{2}$ supported count vectors and taking the best one. The theorem is
exact for the specified probability profile, including profiles with irrational
$t_*$. For the original unweighted finite-dataset question, the rational
approximation argument in Theorem~\ref{thm:ob-main} supplies its precise
consequence: the lower bound $F(k,n)\geq1+B_{k,n}$. No upper bound for arbitrary
mass vectors is inferred from the fixed-profile certificate.
\subsection{An exact finite minimax engine for every symmetric mass profile}
The matching construction above belongs to a larger family that can be solved
exactly without searching over arbitrary Gram matrices. Define, for
$0\leq r\leq n$,
\begin{equation}\label{eq:ob-mode-values}
a_r=\frac{(r/n-t)^2}{t(1-t)},\qquad
b_r=\frac{s_r(k-s_r)}{(k-1)n^2t},\qquad s_r=r\bmod k.
\end{equation}
These are respectively the error ratios of the pure axial and pure transverse
geometries at the balanced rare-count type $r$.
\begin{theorem}[Exact two-mode game and two-type certificate]
\label{thm:ob-engine}
For every $k\geq2$, $n\geq1$, and $0<t<1$,
\begin{align}
\mathfrak C_{k,n}(t)
&=\max_{0\leq\lambda\leq1}\min_{0\leq r\leq n}
\bigl(\lambda a_r+(1-\lambda)b_r\bigr)\label{eq:ob-mode-game}\\
&=\min_{\psi\in\Delta_{n+1}}
\max\left\{\sum_{r=0}^n\psi_r a_r,
\sum_{r=0}^n\psi_r b_r\right\}.
\label{eq:ob-mode-dual}
\end{align}
An optimal $\psi$ can always be supported on at most two values of $r$.
For each supported $r=ak+s$, its count-type lottery gives every rare point
$a$ copies, chooses a uniform $s$-subset to receive one additional copy, and
fills the remaining $n-r$ slots with the anchor. The resulting mixture satisfies
\[
\mathbb E\lVert h-\mu\rVert^2\leq\mathfrak C_{k,n}(t)V
\]
for every geometry with the specified mass vector. Its worst-case guarantee is
exact.
\end{theorem}
\begin{proof}
In Proposition~\ref{prop:ob-symmetry}, reparameterize the nonnegative axial and
transverse contributions by
\[
\lambda=\frac{(1-t)\rho}{(1-t)\rho+1}\in[0,1].
\]
The normalized error at count type $r$ is precisely
$\lambda a_r+(1-\lambda)b_r$, proving \eqref{eq:ob-mode-game}.
For clarity, the duality can be seen in two dimensions. Let
$K=\operatorname{conv}\{(a_r,b_r):0\leq r\leq n\}$ and let
$\gamma=\min_{z\in K}\max\{z_1,z_2\}$. The compact set $K$ is disjoint from
the open lower square $\{z:z_1<\gamma,z_2<\gamma\}$. A separating supporting
line has a normal vector with nonnegative coordinates, because that square is
unbounded in both negative coordinate directions. Normalize its coordinates to
$(\lambda,1-\lambda)$. The separation implies
\[
\min_{z\in K}\bigl(\lambda z_1+(1-\lambda)z_2\bigr)\geq\gamma.
\]
The reverse inequality follows because any convex combination of two coordinates
is at most their maximum. This proves \eqref{eq:ob-mode-dual}.
A minimizing point of $K$ can be chosen on its boundary: an interior minimizer
could be moved a sufficiently small distance in direction $(-1,-1)$ to decrease
the objective. Every boundary point of a planar polygon lies on an edge or is a
vertex, and hence is a convex combination of at most two generating vertices.
If $K$ is a segment or a singleton, the same assertion is immediate. Thus an
optimal mixture uses at most two count types.
For arbitrary geometry, let $W$ and $y$ be as in
Theorem~\ref{thm:ob-fixed-minimax}. The full copies in a balanced count type
cancel the centered rare-point deviations, leaving only a uniform $s$-subset.
Therefore its expected squared error is
\[
(r/n-t)^2\lVert y\rVert^2+
\frac{s(k-s)}{(k-1)n^2}W.
\]
After division by $V=tW+t(1-t)\lVert y\rVert^2$, this is a convex combination
of $a_r$ and $b_r$, with axial weight
$t(1-t)\lVert y\rVert^2/V$. The dual mixture bounds both coordinates by
$\mathfrak C_{k,n}(t)$ and consequently bounds every such geometry. The matching
worst geometry is supplied by Proposition~\ref{prop:ob-symmetry}.
\end{proof}
\begin{corollary}[At most $k$ count candidates]\label{cor:ob-engine-cost}
Within each residue class modulo $k$, retain only a value $r$ in $[0,n]$ nearest
to $nt$. At most $s=\min\{k,n+1\}$ count types remain, without changing
\eqref{eq:ob-mode-game} or \eqref{eq:ob-mode-dual}. For rational $t$,
$\mathfrak C_{k,n}(t)$ is rational and can be computed using $O(s^3)$ exact
scalar-arithmetic operations, apart from the bit costs of the inputs and output.
\end{corollary}
\begin{proof}
Two count totals in the same residue class have the same $b_r$. The one nearest
to $nt$ minimizes $a_r$, and thus has no larger cost against any
$\lambda\in[0,1]$. It dominates every other total in that residue class for the
selector. Removing the dominated types preserves both game values.
For the remaining $s$ types, define the affine functions
$f_r(\lambda)=b_r+\lambda(a_r-b_r)$. Their pointwise minimum is concave and
piecewise affine. A maximum occurs at $\lambda=0$, $\lambda=1$, or an
intersection of two nonidentical functions lying in $[0,1]$. There are $O(s^2)$
such candidates, each evaluated against $s$ functions. Rational $t$ makes all
coefficients and all nonparallel intersection points rational. This gives the
stated exact-arithmetic bound. Finding the representative of a residue class
requires only rounding to the nearest integer in a bounded arithmetic
progression; it does not require scanning all $n+1$ count totals.
\end{proof}
The operation bound concerns the fixed-profile minimax constant and its
count-type certificate. It does not claim that the unrestricted profile
$F(d,n)$ has been computed, and it is not a constant-bit complexity statement.
\begin{corollary}[Exact rules on the number of rare draws]
\label{cor:ob-rare-rules}
Let $\mathcal A\subseteq\{0,\ldots,n\}$ be nonempty and permit exactly those
selections whose total number of non-anchor draws lies in $\mathcal A$.
Within every permitted total, allow all count vectors, including the balanced
ones. Then Theorem~\ref{thm:ob-engine} remains valid with the minimizations and
mixtures restricted to $r\in\mathcal A$. An optimal certificate still uses at
most two permitted balanced count types. Within each residue, only a permitted
$r$ nearest to $nt$ is needed.
\end{corollary}
\begin{proof}
The rule is invariant under permutations of the non-anchor points, so the Gram
symmetrization proof is unchanged. For each permitted total, balancing is still
legal and minimizes transverse error. Restricting the finite two-mode game to
those totals preserves the planar convex-hull proof of duality and the two-type
support bound. The residue dominance argument also applies without change.
\end{proof}
This covers mandatory anchor retention, an upper bound on anchor repetition,
and prescribed rare-count levels. The condition that balanced count vectors are
allowed is part of the theorem. Arbitrary identity-specific restrictions need
not satisfy it and are handled by the general count-certificate formulation
instead. For an explicitly listed set $\mathcal A$, reading and pruning that
list costs an additional $O(|\mathcal A|)$ arithmetic operations.