File size: 16,296 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
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
\section{What the theory permits a learning system to change}
\label{sec:rsi-contract}

The research supplies a concrete mathematical component for recursive
improvement: a learner can change how it retains and replays observations
while preserving a checkable statement about a fixed empirical objective.
The admissible change is not required to be chosen by a fixed algorithm.
A learned policy, a search procedure, or a redesigned selector may propose
the counts. Correctness rests on the returned counts and their certificate.

\subsection{A certificate-based update protocol}

At an anchor state, compute a feature map appropriate to the required
guarantee. Examples are observations themselves, a quadratic loss's Hessian
and residual-gradient pair, or a tuple of derivative tensors. Record the
original observation identities, the feature definition, and the relevant
full-data moments. The meaning of these quantities must remain fixed while
their certificate is used.

The following sequential protocol is directly covered by the preceding
theorems.
\begin{enumerate}
 \item A candidate retention policy chooses an integer budget and legal
 counts. It also supplies either a directly computed feature discrepancy,
 an exact matrix certificate, or a valid upper bound derived from a proved
 selection constant.
 \item The verifier checks count integrality, total budget, original-example
 membership, rule feasibility, and the claimed error inequality against
 the retained original moments.
 \item A learning update uses the certified feature error. For quadratic
 ERM, the joint certificate controls positive definiteness and full-objective
 excess. For gradient updates, an anchor or derivative certificate supplies
 a reuse radius and the observable descent condition.
 \item Before an update leaves that radius, the system refreshes its anchor
 or strengthens its retained feature information. If the representation
 changes, it establishes a new feature certificate from adequate original
 data rather than treating the old one as a guarantee for the new features.
 \item For a sequence of approximate updates, it retains the local errors
 and Lipschitz factors appearing in Theorem~\ref{thm:recursive-error}.
\end{enumerate}

Under the hypotheses of Theorem~\ref{thm:observable-descent}, every accepted
gradient step decreases the same full empirical objective by the stated
amount. Under the additional PL hypothesis, successive accepted steps
satisfy the displayed contraction in objective gap. These are mathematical
guarantees for a defined update process. They do not depend on how the
candidate selection policy was discovered.

This separation permits improvement of the selector itself: a replacement
policy may find a lower-cost or lower-error replay set, and the verifier can
compare its certificate under the unchanged task. Establishing a net speed
gain still requires measuring selection, verification, refresh, memory and
training costs. The theorem does not declare a costly certificate profitable
merely because its replay multiset is small.

\subsection{Why the higher-dimensional results matter to that protocol}

Three mathematical distinctions become operational.
First, intrinsic feature rank determines the available geometric constant.
A high-dimensional neural parameter vector can generate a low-rank gradient
cloud, but that is a property to establish, not an automatic assumption.
Conversely, lifting a quadratic learner to Hessians and residuals may increase
the relevant dimension substantially.

Second, a replay constraint has a quantifiable price. Exact quotas and caps
belong in the admissible count set; positive-margin linear rules can use
Theorem~\ref{thm:margin-rule}; unbiased stochastic estimation has the exact
$1/n$ barrier. These interfaces are mathematically different, and changing
one can invalidate a previously sharp constant.

Third, the rare-simplex obstruction identifies a specific failure mode for
retention. The learner must balance displacement away from a frequent anchor
against transverse diversity among rare points. Optimizing only a rare-point
frequency or only simplex coverage misses their interaction. The exact
two-mode solver provides an auditable benchmark for this interaction over
every symmetric mass profile, while the all-budget obstruction rules out a
universal extrapolation of planar constants.

\subsection{A worked distinction between local and global claims}

The rational example in Section~\ref{sec:quadratic-example} simultaneously
certifies a compressed quadratic solution and a reused gradient step.
Its feature cloud has rank one despite a two-dimensional parameter space,
and the selected examples have different Hessians. Its observed excess
$2/121$ is bounded by a posterior certificate $6/235$ and a universal
budget certificate $2/25$. Those numbers are exact.

Nothing in that example proves a universal acceleration factor for neural
training. What it establishes is that the curvature and replay conditions
can be nonvacuous together, that a certificate can improve on a universal
bound on a concrete instance, and that the same retained data can support
a verified update beyond the initial anchor. The pathwise recursion theorem
states exactly what additional stability information is needed when such
components are composed.

\section{Algorithms, complexity and exact verification}
\label{sec:implementation}

\subsection{The unrestricted selector}

The file \texttt{selection.py} implements the following construction with
\texttt{fractions.Fraction} throughout.
\begin{enumerate}
 \item Process the original data once, applying exact affine elimination
 while storing the original count, first-moment sum and second-moment sum.
 \item On the resident support, compute floor multiplicities and solve its
 fractional residual using the atomic selector.
 \item Check the actual residual selection error against the universal
 coefficient times the resident variance. If it already satisfies the
 universal guarantee, return immediately.
 \item Otherwise use the pair-padding-peeling construction. At large
 budgets repeat a close point if possible; if not, peel a legal unit of
 mass. At small budgets use the exact pair search and deterministic padding.
\end{enumerate}

The residual shortcut preserves the universal guarantee because acceptance
uses an exact inequality. Its remaining noninteger subproblem has fewer
slots than the number of resident points. For a fixed positive-variance
distribution it eventually succeeds at sufficiently large budgets: its
error is $O(n^{-2})$, while the universal allowance is of order $1/n$.
A returned count vector represents a billion replay slots without listing
them individually. Consuming all those slots in an actual learner remains
a separate cost.

If the original data have affine rank at most $r$ and ambient coordinate
dimension $d$, straightforward elimination costs
$O(Nd(r+2)^2)$ exact arithmetic operations. The direct selection stage costs
$O(nd(r+1)+d(r+1)^2)$ in the worst case. Resident streaming observations use
$O(d(r+1))$ scalar registers. The supplied implementation also retains an
optional-to-the-mathematics construction trace and stack of up to $O(n)$
entries. The JSON wrapper loads the input file normally. These are arithmetic
operation and storage-count bounds, not constant-bit or uniform wall-clock
claims.

The output records original row indices and integer multiplicities, the
original mean and variance, resident weights, the attained squared error,
the proved upper coefficient, and the chosen construction method. The
implementation does not label this result instance-optimal.

\subsection{Exact optimization with arbitrary rules}

For small input support, \texttt{selection.py --exact} enumerates all
$\binom{n+m-1}{m-1}$ count vectors. It tests per-observation lower and upper
count bounds and rational linear comparisons, then minimizes the exact
squared mean error among the feasible vectors. It returns either a certified
finite optimum or explicit infeasibility. A configurable state cap prevents
an accidentally enormous enumeration from being presented as a completed
search.

The file \texttt{certificates.py} checks rational covariance upper certificates
and Gram lower certificates. Its lower checker enumerates the full legal
count set independently, including a completeness check when an admissible
list is supplied. Its upper checker needs only the supported lottery counts,
their exact masses, the selection rules, and a rational positive
semidefinite comparison. It deliberately rejects an approximate floating
point inequality as an exact certificate.

\subsection{The symmetric-profile minimax engine}

The file \texttt{profile\_minimax.py} computes the exact fixed-profile value
for masses $(1-t,t/k,\ldots,t/k)$. It reduces the count types by residue,
examines the endpoints and intersections of the remaining affine functions,
and computes a dual mixture independently by checking one- and two-type
mixtures. The returned primal and dual values must agree exactly.

Its output includes the rational constant, the axial fraction of a
normalized centered Gram witness, the complete Gram matrix, at most two
balanced count types, and their lottery probabilities. Expanding a type
means choosing a uniform subset of the appropriate size; the compact
description avoids listing every permutation. The default verification
uses the residue representatives justified by the dominance proof.
The optional \texttt{--verify-all-types} flag checks all $n+1$ rare-count
totals at an additional $O(n)$ cost.

For example,
\begin{verbatim}
python profile_minimax.py --k 5 --n 4 --t 1/2
python profile_minimax.py --k 4 --n 3 --t 3/5
\end{verbatim}
return exact excess values $3/20$ and $61/297$. Their normalized lower Gram
certificates and matrix-equality upper lotteries are included separately.
The former uses 15 supported lottery count vectors and checks all 126 legal
counts; the latter checks all 35 legal counts. These are solved fixed-mass
problems, not claims that $F(5,4)=23/20$ or $F(4,3)=358/297$.

\subsection{What was verified computationally}

The companion reports record exact finite checks and their actual counts.
The principal groups are shown below; they are not added into a single
score, because several groups intentionally inspect overlapping instances.
\begin{center}
\small
\begin{tabular}{p{0.58\textwidth}r}
\toprule
Verification group & Cases\\
\midrule
Original-data selector outputs, dimensions 1--8 & 157\\
Weighted atomic selections & 191\\
Fractional-residual certificates & 191\\
Independent constrained optimum comparisons & 32\\
Exact coefficient identities & 1040\\
Unrestricted rational symmetric profiles & 140\\
Restricted symmetric profiles & 70\\
Generic matrix checks of profile-solver outputs & 11\\
Count vectors for the integer-coordinate obstruction & 35\\
Additional rational-Gram obstruction count vectors & 2763\\
Rational unbiasedness examples & 60\\
Exact recursive equality cases & 60\\
\bottomrule
\end{tabular}
\end{center}

The selector verifier recomputes original variance using the pairwise-distance
identity, rather than the selector's central-moment routine. Its constrained
reference optimizer enumerates Cartesian tuples rather than using the
implementation's count-composition recursion. A forced residual failure
checks the direct fallback, and a billion-slot budget checks compact output.
The profile solver is cross-checked both against its unpruned scalar game
and against generic rational matrix upper and exhaustive Gram lower
certificates. Invalid certificates include tiny exact violations, illegal
counts, omitted admissible vectors, incorrect normalization and singular
indefinite matrices.

The universal mathematical conclusions rest on the proofs. These finite
checks validate algebraic examples, implementation branches and certificates;
they are not proof-assistant formalization, empirical neural-training
benchmarks, or independent scientific replication.

\section{Resolution map and research boundaries}
\label{sec:resolution}

The following table summarizes excess factors, so add one to obtain the
full loss ratio. Brackets denote proved lower and upper bounds for the
unrestricted problem.
\begin{center}
\small
\begin{tabular}{c c c c}
\toprule
Dimension & $n=2$ & $n=3$ & $n=4$\\
\midrule
1 or 2 & $1/3$ & $1/5$ & $1/7$\\[3pt]
3 & $1/3$ & $[1/5,\,7/27]$ & $[1/7,\,21/115]$\\[3pt]
4 & $3/8$ & $[(11-\sqrt{13})/36,\,5/18]$
  & $[(15-\sqrt{33})/64,\,7/32]$\\[3pt]
5 & $2/5$ & $[(16-\sqrt{40})/45,\,13/45]$
  & $[3/20,\,9/40]$\\
\bottomrule
\end{tabular}
\end{center}
The exact pair entries are prior results; the planar all-budget entries
are inherited and re-proved. The symmetric-profile theorems resolve a
separate family for every $k,n,t$ through a finite exact computation.
The unresolved cell $(d,n)=(3,3)$ is deliberately retained as an interval.
Numerical absence of a counterexample would not close that interval.

\begin{figure}[p]
\centering
\includegraphics[width=\textwidth]{figures/selection_bounds.pdf}
\caption{Proved bounds and an asymptotic obstruction. The universal upper
profile and rare-simplex lower family leave a gap; the curves are not an
assertion of a sharp unrestricted profile. The right panel shows a necessary
asymptotic coefficient, not the exact coefficient of $F(d,n)-1$.}
\label{fig:bounds}
\end{figure}

\begin{figure}[p]
\centering
\includegraphics[width=0.94\textwidth]{figures/profile_minimax.pdf}
\caption{Exact fixed-profile values for $k=5,n=4$ at rational sampled masses.
Every plotted sample is computed by exact rational arithmetic; connecting
segments are visual guides. The marked value $\mathfrak C_{5,4}(1/2)=3/20$
has matching matrix and Gram certificates. These are worst-case values over
geometry at the specified masses, not empirical training measurements.}
\label{fig:profile}
\end{figure}

\subsection{Established ingredients and additional deductions}

Affine moment compression, finite-dimensional convex separation, variance
decomposition, Taylor expansion and smoothness-based descent are established
mathematical tools. Exact weighted least-squares coresets and gradient
coreset selection also have substantial prior literature
\cite{maalouf2019,craig2020}. The pair coefficient is explicitly credited
to \cite{pair}. Approximate Carathéodory work concerns related representation
questions \cite{combettes2023,tinarrage2026}; radius and variance
normalizations cannot be exchanged without justification.

The principal deductions developed here are the closed all-budget
rare-simplex obstruction with its optimizer and matching count coupling;
the exact all-mass symmetric-profile reduction and residue-based engine;
the constructive higher-dimensional upper profile and residual allocation
contract; and the joint curvature--residual transfer to equal-weight
quadratic ERM. The richer rule, matrix, derivative, and recursive results
make explicit how such selection contracts can be used together. Some
individual deductions may admit alternate formulations in existing
approximation or optimization literature. A targeted primary-source review
did not locate the same all-budget obstruction or exact symmetric-profile
engine; this is not a certification of worldwide priority.

\subsection{What remains outside the claims}

The manuscript does not determine $F(d,n)$ for every pair $(d,n)$, prove
optimality of its universal upper profile, or prove that the rare-simplex
family contains a global unrestricted extremizer. General count-rule
optimization can be combinatorial; the polynomial scalar engine applies
to the stated symmetric mass and count-type setting. A proof of empirical
loss preservation does not by itself establish distribution-shift robustness,
and the population theorem uses its stated IID finite-variance assumptions.

Recursive improvement of an unrestricted AI is not a theorem of this model.
What is established is a collection of exact selection problems and
certified learning transformations with explicit conditions for feasibility,
descent, reuse and composition. Those conditions make the proposed learning
component mathematically reviewable and give concrete targets for future
algorithmic or experimental work.