File size: 9,851 Bytes
19591f3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
\section{Arbitrary count rules as an exact semidefinite game}
\label{sec:rule-game}

The previous exact profiles exploit symmetry. A general rule system may
instead prescribe a finite set
\[
 \mathcal C\subseteq\{c\in\mathbb Z_{\geq0}^{m}:\mathbf1^\top c=n\}
\]
of admissible integer counts. It may encode lower and upper repetition caps,
exact category quotas, a rational linear cost budget, mandatory observations,
forbidden observations, or a precomputed list of permissible curricula.
We assume $\mathcal C$ is nonempty; otherwise the selection task is infeasible.
These are restrictions on original observation identities. A geometric
compression that discards identities cannot automatically preserve them.

For $m=1$ the sole admissible count vector represents the mean exactly; set
the geometric difficulty to zero. Henceforth assume $m\geq2$.
Fix a strictly positive probability vector $p$ on these $m$ identities, and
put $D_p=\diag(p)$, $s=(\sqrt{p_i})_i$, and
\[
 q_c=c/n,\qquad u_c=D_p^{-1/2}(q_c-p).
\]
Since both $q_c$ and $p$ sum to one, $u_c\perp s$. Define the relative
geometric difficulty of the rules by
\begin{equation}\label{eq:rule-minimax-definition}
 \Gamma(p,\mathcal C)=
 \sup_{(x_i):V_p>0}\min_{c\in\mathcal C}
 \frac{\norm{\sum_i(q_{c,i}-p_i)x_i}^2}{V_p},
\end{equation}
where the supremum allows Hilbert-space dimension up to $m-1$.

\begin{theorem}[Exact primal and dual rule formulations]
\label{thm:rule-minimax}
The supremum in \eqref{eq:rule-minimax-definition} is attained, and
\begin{align}
 \Gamma(p,\mathcal C)
 &=\max_{B\succeq0,\,Bs=0,\,\tr B=1}
             \min_{c\in\mathcal C}u_c^\top Bu_c\label{eq:rule-primal}\\
 &=\min_{\lambda\in\Delta(\mathcal C)}
       \lambda_{\max}\left(\left.
       \sum_c\lambda_cu_cu_c^\top\right|_{s^\perp}\right).
       \label{eq:rule-dual}
\end{align}
An optimal dual lottery may be supported on at most
$m(m-1)/2+1$ count vectors. It depends on the masses and admissible count set,
and supplies the same upper bound for every support geometry.
\end{theorem}
\begin{proof}
Center a point configuration so that $\sum_i p_i x_i=0$, and let
$G_{ij}=\ip{x_i}{x_j}$. Then $G\succeq0$, $Gp=0$, and
$V_p=\tr(D_pG)$. After scaling to unit variance, the congruence
$B=D_p^{1/2}GD_p^{1/2}$ is positive semidefinite, satisfies $Bs=0$,
and has trace one. Conversely any such $B$ produces
$G=D_p^{-1/2}BD_p^{-1/2}$ with those properties, which can be factored as a
Gram matrix in dimension at most $m-1$. Its selected error is
$u_c^\top Bu_c$. The feasible set of $B$ is compact, proving
\eqref{eq:rule-primal} and attainment.

We prove the dual identity directly by convex separation. Work on
$s^\perp$, which has dimension $m-1$, and let
$\mathcal K=\conv\{u_cu_c^\top:c\in\mathcal C\}$.
For any $M\in\mathcal K$ and positive semidefinite trace-one $B$,
\[
 \min_c\tr(Bu_cu_c^\top)\leq\tr(BM)\leq\lambda_{\max}(M).
\]
Thus the primal value $v$ does not exceed the dual value. Suppose, for some
$a>v$, no $M\in\mathcal K$ satisfies $M\preceq aI$.
The compact set $\mathcal K$ and the closed convex set
$\{M:M\preceq aI\}$ are strictly separable. The separating functional can
be written $M\mapsto\tr(HM)$. Its supremum on the second set is finite only
if $H\succeq0$, and then equals $a\tr H$. Strict separation implies
\[
 \min_c\tr(Hu_cu_c^\top)>a\tr H.
\]
Here $H\ne0$, so normalization by its positive trace gives an admissible
primal matrix with objective greater than $a>v$, a contradiction. Therefore
the dual value is at most every $a>v$ and hence at most $v$. Compactness
of the lottery simplex gives an optimizer. Finally the vector space of
symmetric matrices on $s^\perp$ has dimension $m(m-1)/2$; affine elimination
preserves the optimal matrix using at most one more lottery atom.
\end{proof}

For a fixed smaller ambient dimension $d<m-1$, the primal must additionally
require $\rank B\leq d$. That rank constraint is nonconvex. The unrestricted
dual is still a valid upper bound, but equality for the rank-constrained
problem is not asserted. The exact symmetric-profile theorem above avoids
this issue because its maximizing geometry is realized in the indicated
dimension $k$.

\subsection{Rational certificates without square roots}

The normalized formulation is convenient for proofs. An equivalent upper
certificate can be checked using rational arithmetic on the original
counts and masses.

\begin{proposition}[A checkable matrix upper certificate]
\label{prop:rational-upper}
Suppose $\lambda_c\geq0$ sums to one on admissible counts and, for $C\geq0$,
\begin{equation}\label{eq:rational-matrix-certificate}
 M:=\sum_c\lambda_c(q_c-p)(q_c-p)^\top
 \preceq C\bigl(D_p-pp^\top\bigr).
\end{equation}
Then every geometry satisfies
\[
 \E_\lambda\norm{h_c-\mu_p}^2\leq CV_p.
\]
The least possible $C$ in \eqref{eq:rational-matrix-certificate} is exactly
$\Gamma(p,\mathcal C)$.
\end{proposition}
\begin{proof}
Let $X$ have the support vectors as columns. Congruence of
\eqref{eq:rational-matrix-certificate} by $X$, followed by a trace, gives
the claimed expectation. For equivalence with the dual, congruence by
$D_p^{-1/2}$ maps $D_p-pp^\top$ to $I-ss^\top$, the orthogonal projector
onto $s^\perp$, and maps $M$ to $\sum_c\lambda_cu_cu_c^\top$. Both matrices
annihilate $s$, so their positive semidefinite comparison is precisely
the eigenvalue inequality in \eqref{eq:rule-dual}.
\end{proof}

A matching lower certificate is a rational matrix $G\succeq0$ with
$Gp=0$, $\tr(D_pG)=1$, and
\[
 q_c^\top Gq_c\geq C\qquad\text{for every }c\in\mathcal C.
\]
Factoring $G$ realizes a unit-variance geometry, proving
$\Gamma\geq C$. A lower certificate must cover the entire admissible count
set; checking only the count vectors used by an upper lottery is insufficient.
When both certificates use the same value, the fixed-mass rule problem is
solved exactly.

Positive semidefiniteness of a rational symmetric matrix can be checked by
exact Schur complements. A negative diagonal entry rejects the matrix. A
zero diagonal must have an identically zero row and column in a positive
semidefinite matrix; otherwise a two-coordinate quadratic form is negative
for one sign and a sufficiently small magnitude. A positive diagonal pivot
reduces the problem to its rational Schur complement. Repeating this process
ends at a zero matrix or a rejection. No numerical eigenvalue tolerance is
required by the supplied certificate checker.

\subsection{Why a new rule may have a real statistical price}

\begin{proposition}[A geometric obstruction from infeasible target masses]
\label{prop:rule-separation}
If $p\notin\conv\{q_c:c\in\mathcal C\}$, there exist $a\in\R^m$ and
$\delta>0$ such that, after reversing $a$ if needed,
\[
 a^\top(q_c-p)\geq\delta\quad\text{for every }c\in\mathcal C.
\]
For the one-dimensional geometry $x_i=a_i$, whose variance is
$V_a=\sum_i p_i(a_i-a^\top p)^2>0$, every legal selection has relative
squared error at least $\delta^2/V_a$. Moreover, a count-level unbiased
lottery, meaning $\E q_c=p$ and hence mean-unbiased for every geometry,
exists if and only if
$p\in\conv\{q_c:c\in\mathcal C\}$.
\end{proposition}
\begin{proof}
The finite convex hull is compact; strict separation gives $a,\delta$.
The vector $a$ cannot be constant because every $q_c-p$ has sum zero.
Positive $p$ then gives $V_a>0$. Its selected mean differs from the full
mean by $a^\top(q_c-p)$, proving the lower bound. The unbiasedness statement
is exactly the definition of a convex representation of $p$ by the $q_c$.
\end{proof}

For example, a prescribed exact category frequency $\alpha$ is arithmetically
impossible unless $n\alpha$ is an integer. A requirement to use distinct
row identities means $c_i\leq1$, which changes the admissible set and can
make a large budget infeasible. The universal upper profile for unrestricted
repetition therefore cannot simply be attached to a constrained output.
The exact small-instance optimizer in the package reports infeasibility
explicitly and enumerates the original identities without lossy compression.

\subsection{Stratified replay and multiple preserved quantities}

Partition observations into groups $g=1,\ldots,G$ with target masses
$\alpha_g>0$, group means $\mu_g$ and within-group variances $V_g$.
If the student must use exactly $n_g=n\alpha_g$ examples in group $g$,
assume each $n_g$ is a positive integer. A groupwise selector provides means
$h_g$ with squared errors at most $c_gV_g$. The overall mean obeys
\begin{equation}\label{eq:stratified-bound}
 \norm{\sum_g\alpha_gh_g-\sum_g\alpha_g\mu_g}^2
 \leq\left(\sum_g\alpha_g\sqrt{c_gV_g}\right)^2.
\end{equation}
This is the triangle inequality applied to the group errors. If their
error subspaces are orthogonal, the stronger bound is
$\sum_g\alpha_g^2c_gV_g$. Biased selections do not generally have canceling
cross terms, so that stronger formula needs the stated orthogonality.

For weighted teacher memory, extra linear moments can be preserved exactly.
Let columns $a_i$ of a matrix $A$ include a row of ones and all features to
preserve. For any feasible probability vector, affine elimination inside
$\ker A$ while not increasing a chosen linear cost produces a feasible
representation with at most $\rank A$ positive weights. The proof is the
same signed-direction argument as Lemma~\ref{lem:compression-general};
every step leaves $Ap$ unchanged. Adding category indicators or quadratic
feature entries increases the possible support requirement. This statement
concerns weighted moment feasibility; integer replay counts require their
own rule certificate.

The next sections show how a scalar geometric guarantee can instead enforce
many soft or margin-based conditions directly in the original feature rank,
and how the same machinery extends from mean prediction to learning rules
whose curvature and gradients change across examples.