File size: 30,623 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 | % Intermediate manuscript-ready sections. All results are self-contained.
% Required macros: \R, \E, \norm{...}, \ip{...}{...}, \tr.
% Theorem environments: theorem, lemma, proposition, corollary, proof.
% Bibliography keys to add: craig2020, maalouf2019, karimi2016.
\section{A common interface for richer learning rules}
\label{sec:learning-interface}
Fix an integer budget $n$ and an affine-rank bound $r$. Throughout the next
sections, $c=c_r(n)>0$ denotes any established universal selection constant:
every finitely supported probability distribution $(p_i,x_i)$ of affine rank
at most $r$ has an integer count vector $k\geq0$, $\sum_i k_i=n$, satisfying
\begin{equation}\label{eq:generic-selection-contract}
\left\|\sum_i\frac{k_i}{n}x_i-\sum_i p_i x_i\right\|^2
\leq c\sum_i p_i\left\|x_i-\sum_j p_jx_j\right\|^2.
\end{equation}
The same $c$ is assumed valid after any fixed linear transformation of the
points, since that transformation cannot increase their affine rank. The
higher-dimensional bounds proved earlier, or the exact planar value
$c_2(n)=1/(2n-1)$, can be substituted without changing any proof below.
These sections do not presuppose an exact answer to the unrestricted
higher-dimensional minimax problem.
Write $\mu=\sum_i p_i x_i$, $\Sigma=\sum_i p_i(x_i-\mu)(x_i-\mu)^\top$,
and $h_k=\sum_i k_i x_i/n$. For a positive semidefinite matrix $Q$, applying
\eqref{eq:generic-selection-contract} to $Q^{1/2}x_i$ gives a selection with
\begin{equation}\label{eq:metric-contract}
(h_k-\mu)^\top Q(h_k-\mu)\leq c\tr(Q\Sigma).
\end{equation}
The selected counts may depend on $Q$. This quantifier is central: a selector
optimal for one metric need not satisfy another metric with the same factor.
\subsection{One random selection policy for every quadratic metric}
\begin{theorem}[Matrix lifting of a scalar selection theorem]
\label{thm:matrix-lottery}
There is a probability distribution $\pi$ over integer count vectors of total
mass $n$ such that
\begin{equation}\label{eq:matrix-lottery}
\E_{k\sim\pi}(h_k-\mu)(h_k-\mu)^\top\preceq c\Sigma.
\end{equation}
If $r_0=\operatorname{rank}\Sigma$, this distribution can be supported on at
most $r_0(r_0+1)/2+1$ count vectors. Consequently, the same policy satisfies
\begin{equation}\label{eq:lottery-metrics}
\E_{k\sim\pi}(h_k-\mu)^\top Q(h_k-\mu)
\leq c\tr(Q\Sigma)
\end{equation}
for every fixed $Q\succeq0$.
\end{theorem}
\begin{proof}
When $\Sigma=0$, every supported point is $\mu$, and every count vector works.
Otherwise work on the $r_0$-dimensional range of $\Sigma$, where $\Sigma$ is
positive definite. There are finitely many count vectors. Let
\[
\mathcal K=\operatorname{conv}\{(h_k-\mu)(h_k-\mu)^\top:k\geq0,
\ \sum_i k_i=n\},\qquad
\mathcal D=\{M:M\preceq c\Sigma\}.
\]
Suppose these convex sets are disjoint. The first is compact and the second is
closed, so a strict separating linear functional exists. Such a functional has
the form $M\mapsto\tr(QM)$ with $Q$ symmetric. Its supremum on $\mathcal D$ is
finite only when $Q\succeq0$: if $u^\top Qu<0$, the matrices
$c\Sigma-tuu^\top$ belong to $\mathcal D$ and make that supremum infinite.
For $Q\succeq0$, the supremum is $c\tr(Q\Sigma)$. Separation would therefore
give
\[
\min_k(h_k-\mu)^\top Q(h_k-\mu)>c\tr(Q\Sigma),
\]
contradicting \eqref{eq:metric-contract}. Thus $\mathcal K\cap\mathcal D$ is
nonempty. A convex representation of an element in the intersection gives
$\pi$ and \eqref{eq:matrix-lottery}.
The real vector space of symmetric $r_0\times r_0$ matrices has dimension
$s=r_0(r_0+1)/2$. If a convex representation uses more than $s+1$ matrices,
their augmented coordinate vectors have an affine dependence. Moving the
weights in that dependence until one vanishes preserves the represented
matrix. Repetition leaves at most $s+1$ positive weights. Taking the trace
against $Q$ proves \eqref{eq:lottery-metrics}.
\end{proof}
This is an existence theorem obtained from finite-dimensional convex
separation; it does not provide a polynomial-time algorithm for finding the
lottery. Enumerating all count vectors can be expensive. Its value is a precise
change in the guarantee: an entire family of quadratic evaluation rules can
share one selection policy. The matrix on the left includes both conditional
variance and the outer product of conditional bias. In particular, the theorem
does not assert $\E h_k=\mu$.
The metric in \eqref{eq:lottery-metrics} must be fixed before the lottery draw,
or chosen from past information independently of that fresh draw. A metric
chosen after seeing the realized error may concentrate on its direction, and
is not covered by the expected guarantee.
\begin{proposition}[A deterministic guarantee for all metrics]
\label{prop:deterministic-matrix}
There is a single count vector with
\begin{equation}\label{eq:deterministic-matrix}
(h_k-\mu)(h_k-\mu)^\top\preceq r_0c\Sigma.
\end{equation}
The rank factor cannot universally be removed: for $n=1$ and the uniform
distribution on the vertices of a regular simplex in $\R^{r_0}$, every
deterministic selection requires factor $r_0$, whereas the scalar constant is
$c_{r_0}(1)=1$.
\end{proposition}
\begin{proof}
Whiten on the covariance range. Equation~\eqref{eq:metric-contract}, with
$Q=\Sigma^+$, gives $e^\top\Sigma^+e\leq cr_0$, where $e=h_k-\mu$ and $+$
denotes the Moore--Penrose inverse. For every vector $v$, Cauchy--Schwarz on
this range yields
$(v^\top e)^2\leq(e^\top\Sigma^+e)(v^\top\Sigma v)$, which is exactly
\eqref{eq:deterministic-matrix}. For unit simplex vertices centered at zero,
$\Sigma=I/r_0$. Every one-point error has norm one, so $ee^\top\preceq K\Sigma$
requires $K\geq r_0$ by testing the direction $e$.
\end{proof}
For $K$ prescribed metrics $Q_j$ with $V_j=\tr(Q_j\Sigma)>0$, a simpler
deterministic scalarization uses $Q=\sum_j Q_j/V_j$. It gives
$\sum_j e^\top Q_je/V_j\leq Kc$, hence every normalized objective is at most
$Kc$. The lottery achieves $c$ for each objective in expectation. These are
different interfaces, not interchangeable statements.
\subsection{Linear rules and an explicit price for feasibility}
\begin{theorem}[Variance-normalized margin rule]
\label{thm:margin-rule}
Suppose a proposed replay mean must satisfy $a_j^\top h\leq b_j$ for
$j=1,\ldots,K$. Assume the full mean has positive margins
$s_j=b_j-a_j^\top\mu>0$. Choose $B\succeq0$ such that
\begin{equation}\label{eq:rule-ellipsoid}
B\succeq \frac{a_ja_j^\top}{s_j^2}\qquad(j=1,\ldots,K).
\end{equation}
For an objective metric $Q_0\succeq0$, write
$S=\tr(B\Sigma)$ and $V_0=\tr(Q_0\Sigma)$. If $cS<1$, there exists an
equal-weight selection of $n$ original examples satisfying every rule and
\begin{equation}\label{eq:margin-penalty}
(h-\mu)^\top Q_0(h-\mu)\leq\frac{cV_0}{1-cS}.
\end{equation}
The affine-rank parameter remains at most $r$ regardless of the number of
linear rules.
\end{theorem}
\begin{proof}
First assume $V_0>0$ and put $t=cV_0/(1-cS)>0$. Apply
\eqref{eq:metric-contract} to $Q_0+tB$. Its right side is
$c(V_0+tS)=t$, so the resulting error satisfies
\[
e^\top Q_0e+t e^\top Be\leq t.
\]
This proves \eqref{eq:margin-penalty} and $e^\top Be\leq1$.
Equation~\eqref{eq:rule-ellipsoid} gives $|a_j^\top e|\leq s_j$, whence
$a_j^\top h\leq a_j^\top\mu+s_j=b_j$ for every $j$.
If $V_0=0$, each supported displacement lies in the nullspace of $Q_0$,
so every selection has zero $Q_0$ error. Selecting in metric $B$ gives
$e^\top Be\leq cS<1$, and the same feasibility argument applies.
Finally $Q_0+tB$ acts linearly on the original cloud, so it does not increase
its affine rank.
\end{proof}
A directly available choice is
$B=\sum_j a_ja_j^\top/s_j^2$. A smaller value of $S$ can be sought by the
semidefinite program minimizing $\tr(B\Sigma)$ subject to
\eqref{eq:rule-ellipsoid}. Thus correlated or redundant rules need not pay
their variances repeatedly. The theorem is a sufficient condition, not a
characterization of all feasible integer selections. Exact equalities have
zero margin and require separate arithmetic feasibility: for example, an
exact group proportion $\alpha$ in an equal-weight sample requires
$n\alpha\in\mathbb Z$. An irrational target proportion cannot be represented
by any finite equal-weight sample.
\section{A sharp barrier imposed by unbiasedness}
\label{sec:unbiased}
Adding an apparently benign rule can change the optimal mathematical law.
Consider selectors that may randomize but must satisfy the conditional rule
$\E[h\mid D]=\mu_D$ for every input dataset. This is stronger than the matrix
lottery theorem, which deliberately permits bias.
\begin{theorem}[Exact minimax cost of the unbiasedness rule]
\label{thm:unbiased-barrier}
For every $d\geq1$ and $n\geq1$, the sharp universal expected squared-error
factor among conditionally unbiased equal-weight selectors is $1/n$:
\begin{equation}\label{eq:unbiased-value}
\inf_{\mathcal A:\E[h\mid D]=\mu_D}
\sup_{D:V_D>0}\frac{\E[\norm{h-\mu_D}^2\mid D]}{V_D}
=\frac1n.
\end{equation}
The output consists of $n$ original examples with repetition. Consequently,
any universal improvement below $1/n$ requires a selector that is biased on
some datasets.
\end{theorem}
\begin{proof}
Independent sampling from the empirical distribution is conditionally
unbiased, and independence makes its expected squared error exactly $V_D/n$.
This proves the upper bound.
For the lower bound, consider a dataset on $\{0,1\}\subset\R$ with a fraction
$p>0$ of ones. Any permitted output mean belongs to
$\{0,1/n,\ldots,1\}$. On this grid, $h^2\geq h/n$. Conditional unbiasedness
therefore implies
\[
\E(h-p)^2=\E h^2-p^2\geq \frac pn-p^2.
\]
The dataset variance is $p(1-p)$, so the ratio is at least
$(1/n-p)/(1-p)$. Taking datasets with one one and $N-1$ zeros, and sending
$N\to\infty$, makes $p=1/N\to0$ and proves the lower bound $1/n$.
The example embeds in every higher dimension. It also respects $n\leq N$
once $N$ is sufficiently large.
\end{proof}
This separates two useful objectives. A selector may minimize the actual
quadratic loss on the observed population, where a controlled bias can be
beneficial, or it may supply an unbiased stochastic estimator for an optimizer.
The sharp planar constant improves the former objective. Theorem~
\ref{thm:unbiased-barrier} prevents presenting that improvement as a free
variance reduction for every unbiased gradient estimator.
\section{From selected means to variable-Hessian quadratic learning}
\label{sec:quadratic-erm}
Let original examples specify convex quadratic losses
\[
f_i(\theta)=\tfrac12\theta^\top H_i\theta-b_i^\top\theta+d_i,
\qquad H_i\succeq0,
\]
and let $F=\sum_i p_if_i$. Assume $H=\sum_i p_iH_i\succ0$ and put
$b=\sum_i p_ib_i$, $\theta_*=H^{-1}b$. The student now minimizes the equal-weight
selected objective $\widehat F=\sum_i k_if_i/n$. This is a richer learner than
the empirical-mean learner; the Hessian may vary across examples.
Use the whitened per-example matrices and residual gradients
\begin{equation}\label{eq:quadratic-lift}
A_i=H^{-1/2}H_iH^{-1/2},\qquad
r_i=H^{-1/2}(H_i\theta_*-b_i).
\end{equation}
Their means are $I$ and $0$. Define
\[
M=\sum_i p_i\norm{A_i-I}_F^2,\qquad
G=\sum_i p_i\norm{r_i}^2.
\]
Here $M$ measures variation of curvature after normalization by the full
objective, and $G$ measures disagreement among example gradients at its
optimizer. The feature pair $(A_i,r_i)$ has affine rank at most
$d(d+1)/2+d$, often less. The constant $c$ in this section is valid for that
feature rank, rather than automatically being the constant for the original
parameter dimension $d$.
\begin{lemma}[Sharp joint perturbation certificate]
\label{lem:joint-perturbation}
Write $\widehat A=\sum_i k_iA_i/n$ and
$\widehat r=\sum_i k_ir_i/n$. If, for some $\tau>0$ and $0\leq E<\tau^2$,
\begin{equation}\label{eq:joint-perturbation-condition}
\tau^2\norm{\widehat A-I}_F^2+\norm{\widehat r}^2\leq E,
\end{equation}
then $\widehat F$ has a unique minimizer $\widehat\theta$ and
\begin{equation}\label{eq:joint-perturbation-bound}
F(\widehat\theta)-F(\theta_*)
\leq\frac{E}{2(1-E/\tau^2)}.
\end{equation}
The final scalar inequality is sharp using only the information in
\eqref{eq:joint-perturbation-condition}.
\end{lemma}
\begin{proof}
Let $u=\norm{\widehat A-I}_{\mathrm{op}}$.
Then $u\leq\sqrt E/\tau<1$, so $\widehat A\succ0$ and the selected Hessian
$\widehat H=H^{1/2}\widehat A H^{1/2}$ is positive definite. The normal
equations give
\[
H^{1/2}(\widehat\theta-\theta_*)
=-\widehat A^{-1}\widehat r.
\]
The exact quadratic excess is therefore
\[
F(\widehat\theta)-F(\theta_*)
=\tfrac12\norm{\widehat A^{-1}\widehat r}^2
\leq\frac{E-\tau^2u^2}{2(1-u)^2}.
\]
The numerator bound uses
$\norm{\widehat A-I}_F\geq u$. For
$0\leq u\leq\sqrt E/\tau$, the derivative of the last fraction has the sign
of $E-\tau^2u$. Its maximum occurs at $u=E/\tau^2$, and substitution gives
\eqref{eq:joint-perturbation-bound}.
For sharpness of this perturbation implication, take one dimension,
$\widehat A=1-E/\tau^2$ and
$\widehat r=\sqrt{E(1-E/\tau^2)}$. The certificate holds with equality, and
so does its conclusion. This is sharpness of the certificate calculation;
it does not assert a minimax lower bound for all quadratic teaching tasks.
\end{proof}
\begin{theorem}[Selected quadratic ERM with optimized curvature weighting]
\label{thm:quadratic-erm}
If $cM<1$, an equal-weight selection of $n$ original examples has a unique
selected minimizer satisfying
\begin{equation}\label{eq:quadratic-main}
\boxed{\displaystyle
F(\widehat\theta)-F(\theta_*)
\leq\frac{cG}{2\bigl(1-\sqrt{cM}\bigr)^2}.}
\end{equation}
When $M=0$, the bound is $cG/2$. When $G=0$, zero excess is achievable under
the same condition $cM<1$.
\end{theorem}
\begin{proof}
For any $\tau>0$, apply the universal selection theorem to the Hilbert-space
features $(\tau A_i,r_i)$, using Frobenius norm on the matrix coordinate.
The selected counts satisfy
\[
\tau^2\norm{\widehat A-I}_F^2+\norm{\widehat r}^2
\leq c(\tau^2M+G)=:E_\tau.
\]
First suppose $M,G>0$. Set $a=cM\in(0,1)$ and $b=cG>0$, and choose
\[
\tau^2=\frac{b}{\sqrt a(1-\sqrt a)}.
\]
Then $E_\tau/\tau^2=\sqrt a<1$ and
$E_\tau=b/(1-\sqrt a)$. Lemma~\ref{lem:joint-perturbation} gives
$b/[2(1-\sqrt a)^2]$, as claimed. This is the minimum of the lemma's bound
over $\tau$: differentiating
$(a t+b)/(1-a-b/t)$ for $t=\tau^2>b/(1-a)$ gives the displayed optimizer.
If $M=0$, every $A_i=I$, so every selected Hessian is exactly normalized to
$I$. Select residuals with $\norm{\widehat r}^2\leq cG$ and use the exact
excess identity in the lemma. If $G=0$, every $r_i=0$. Select matrices with
$\norm{\widehat A-I}_F^2\leq cM<1$; then $\widehat A\succ0$ and the selected
normal equations have solution $\theta_*$ exactly.
\end{proof}
For least squares, $H_i$ and $b_i$ are the per-example design outer product
and response-weighted design vector. A fixed ridge term may be incorporated
in every $H_i$. Thus the theorem applies directly to a fixed equal-weight
ridge-regression learner. It supplies a joint bound on curvature distortion
and gradient residual instead of treating the two perturbations independently.
The teacher needs the full $H$, $b$, and $\theta_*$ to form the normalized
features. The result is a theorem about a constrained representation of that
solution, not a claim that those initial computations disappear.
\subsection{Exact weighted summaries and their established provenance}
The vector consisting of the distinct entries of $H_i$ and the entries of
$b_i$ has dimension $D=d(d+1)/2+d$. Affine elimination therefore yields a
weighted subset of at most $D+1$ original examples preserving both $H$ and
$b$ exactly. It preserves the gradient for every $\theta$ and changes the
objective only by a constant. Preserving $d_i$ as well uses at most $D+2$
points and preserves the complete objective. Any deterministic optimization
procedure depending only on those gradients, and initialized identically,
then follows the same trajectory on the exact weighted summary.
These are classical finite-dimensional moment consequences of
Carath\'eodory-type coreset constructions. Exact least-squares and matrix
coresets are established research topics \cite{maalouf2019}; they are not
claimed here as a new discovery. The equal-weight realization theorem
\ref{thm:quadratic-erm}, its optimized joint certificate, and the explicit
connection to a proved count-budget constant are the additional deductions
developed in this manuscript.
\section{Gradient replay that remains valid inside a trust region}
\label{sec:gradient-reuse}
For differentiable example losses, let $g_i(\theta)=\nabla f_i(\theta)$,
$g(\theta)=\sum_i p_i g_i(\theta)$, and
$\widehat g(\theta)=\sum_i q_i g_i(\theta)$, where $q_i=k_i/n$.
At an anchor $\theta_0$, the gradient cloud has variance
$V_g=\sum_i p_i\norm{g_i(\theta_0)-g(\theta_0)}^2$. Selecting this cloud yields
\begin{equation}\label{eq:anchor-error}
\norm{\widehat g(\theta_0)-g(\theta_0)}\leq\sqrt{cV_g}=:\epsilon_0.
\end{equation}
The affine-rank parameter here is the rank of the gradient cloud at the
anchor. Gradient coreset selection itself is an established approach; for
example, CRAIG selects weighted subsets approximating the full gradient
\cite{craig2020}. The following results state what the present equal-weight
certificate implies, including how long a retained selection remains valid.
\begin{theorem}[A certified radius for reuse]
\label{thm:gradient-radius}
Suppose each $g_i$ is $L_i$-Lipschitz on a convex region containing
$\theta_0$ and $\theta$. Put
$\Lambda=\sum_i|q_i-p_i|L_i$. Then the same count vector satisfies
\begin{equation}\label{eq:gradient-radius}
\norm{\widehat g(\theta)-g(\theta)}
\leq\epsilon_0+\Lambda\norm{\theta-\theta_0}.
\end{equation}
It is sufficient to replace $\Lambda$ by the computable upper bound
$\sum_i q_iL_i+\sum_i p_iL_i$.
\end{theorem}
\begin{proof}
Subtract the error at $\theta_0$ from the error at $\theta$ and use
\[
\widehat g(\theta)-g(\theta)
-\widehat g(\theta_0)+g(\theta_0)
=\sum_i(q_i-p_i)\bigl(g_i(\theta)-g_i(\theta_0)\bigr).
\]
The triangle inequality and the Lipschitz assumptions prove
\eqref{eq:gradient-radius}. The stated upper bound follows from
$|q_i-p_i|\leq q_i+p_i$.
\end{proof}
The term $\sum_i p_iL_i$ is a scalar that can be retained during the initial
scan; the other term uses only selected examples. A tolerance $\epsilon$ is
therefore certified throughout the radius
$(\epsilon-\epsilon_0)/\Lambda$ whenever $\epsilon\geq\epsilon_0$ and
$\Lambda>0$. This controls replay reuse without silently treating a
single-point gradient certificate as uniform over parameter space.
\subsection{Higher-order derivative rules}
The radius bound can be strengthened when a retained selection also
approximates curvature or higher derivatives. Fix an integer $k\geq1$.
Assume each gradient map is $C^k$ on the relevant
ball and $\norm{D^k g_i(\theta)}_{\mathrm{op}}\leq M_i$ there.
For $s=0,\ldots,k-1$, define derivative discrepancies
\[
E_s=\sum_i(q_i-p_i)D^s g_i(\theta_0).
\]
Tensor norms below are the Euclidean coefficient norms, so contracting a
tensor against $s$ copies of $u$ has norm at most its tensor norm times
$\norm u^s$.
\begin{theorem}[Joint derivative certificate]
\label{thm:jet-certificate}
Let $w_s>0$. Suppose the count vector obeys
\[
\sum_{s=0}^{k-1}w_s^2\norm{E_s}_F^2\leq E_{\mathrm{jet}}.
\]
For $\norm{\theta-\theta_0}\leq R$,
\begin{equation}\label{eq:jet-radius}
\norm{\widehat g(\theta)-g(\theta)}
\leq
\sqrt{E_{\mathrm{jet}}}
\left(\sum_{s=0}^{k-1}\frac{R^{2s}}{(s!)^2w_s^2}\right)^{1/2}
+\frac{R^k}{k!}\sum_i|q_i-p_i|M_i.
\end{equation}
Such a certificate follows from the mean-selection theorem by applying it
to the fixed feature vector
$(w_sD^sg_i(\theta_0))_{s=0}^{k-1}$, with
$E_{\mathrm{jet}}=cV_{\mathrm{jet}}$ and the corresponding feature-rank bound.
\end{theorem}
\begin{proof}
Taylor's formula along the segment $\theta_0+t(\theta-\theta_0)$ gives the
gradient error as
\[
\sum_{s=0}^{k-1}\frac1{s!}E_s[(\theta-\theta_0)^{\otimes s}]
+\mathcal R,
\qquad
\norm{\mathcal R}\leq\frac{R^k}{k!}
\sum_i|q_i-p_i|M_i.
\]
The polynomial part has norm at most
$\sum_s\norm{E_s}_F R^s/s!$. Applying Cauchy--Schwarz with factors
$w_s\norm{E_s}_F$ and $R^s/(s!w_s)$ proves
\eqref{eq:jet-radius}. The final assertion is
\eqref{eq:generic-selection-contract} in the direct sum of the derivative
tensor spaces.
\end{proof}
For quadratic losses, retaining gradient and Hessian discrepancies corresponds
to $k=2$ with zero remainder. For more general losses, higher-order rules can
make the residual drift scale as $R^k$, while their feature dimension and
evaluation cost grow. The theorem makes that tradeoff explicit; it does not
assume high-order tensors can be computed cheaply.
\section{Certified optimization and recursive composition}
\label{sec:recursive-composition}
\subsection{Descent using an observable error bound}
Assume the full objective $F$ is $L$-smooth on a region containing a proposed
step. Suppose a replay certificate supplies
$\norm{\widehat g(\theta)-\nabla F(\theta)}\leq\epsilon$.
The error bound may come from the anchor, radius, or derivative theorem above.
\begin{theorem}[Descent from an observable replay certificate]
\label{thm:observable-descent}
If $\widehat g\ne0$ and
$a=\epsilon/\norm{\widehat g}<1$, choose
\[
\eta=\frac{1-a}{L},\qquad \theta^+=\theta-\eta\widehat g.
\]
Provided the step stays inside the region on which smoothness and the error
certificate hold,
\begin{equation}\label{eq:observable-descent}
F(\theta^+)\leq F(\theta)
-\frac{(\norm{\widehat g}-\epsilon)^2}{2L}.
\end{equation}
If additionally $F$ obeys the Polyak--\L{}ojasiewicz inequality
$\norm{\nabla F(\theta)}^2\geq2\mu(F(\theta)-F_*)$, then
\begin{equation}\label{eq:observable-pl}
F(\theta^+)-F_*
\leq\left[1-\frac\mu L
\left(\frac{1-a}{1+a}\right)^2\right](F(\theta)-F_*).
\end{equation}
\end{theorem}
\begin{proof}
Let $e=\widehat g-\nabla F(\theta)$. Smoothness gives
\[
F(\theta-\eta\widehat g)
\leq F(\theta)-\eta\ip{\nabla F(\theta)}{\widehat g}
+\frac{L\eta^2}{2}\norm{\widehat g}^2.
\]
Since $\ip{\nabla F(\theta)}{\widehat g}
\geq\norm{\widehat g}^2-\epsilon\norm{\widehat g}
=(1-a)\norm{\widehat g}^2$, the chosen $\eta$ proves
\eqref{eq:observable-descent}. Also
$\norm{\nabla F(\theta)}\leq(1+a)\norm{\widehat g}$, so the decrease is at
least
$\frac1{2L}((1-a)/(1+a))^2\norm{\nabla F(\theta)}^2$.
The PL inequality proves \eqref{eq:observable-pl}.
\end{proof}
This test uses the retained-data gradient and a valid error radius; it does
not need a fresh exact full gradient merely to evaluate the inequality
$\epsilon<\norm{\widehat g}$. If the test fails, the theorem does not certify
the proposed step. A new anchor, a larger replay budget, or a stronger
derivative summary may restore a certificate.
The smoothness descent argument and PL convergence principle are standard
optimization tools \cite{karimi2016}. Their role here is to turn a proved
selection error, with explicit count-budget dependence, into an observable
update condition. No claim is made that all neural-network objectives satisfy
the PL condition globally.
\begin{corollary}[Nonconvex stationarity from successive certificates]
\label{cor:stationarity}
Suppose $F\geq F_{\inf}$ and the preceding theorem applies for $T$ successive
steps, with $a_t\leq\overline a<1$. Then
\[
\min_{0\leq t<T}\norm{\nabla F(\theta_t)}^2
\leq\frac{2L(F(\theta_0)-F_{\inf})}{T}
\left(\frac{1+\overline a}{1-\overline a}\right)^2.
\]
\end{corollary}
\begin{proof}
The proof of Theorem~\ref{thm:observable-descent} bounds the decrease at step
$t$ below by
$((1-\overline a)/(1+\overline a))^2
\norm{\nabla F(\theta_t)}^2/(2L)$. Sum these inequalities and use the lower
bound on $F$; the smallest gradient norm is no greater than their average.
\end{proof}
\subsection{A pathwise theorem for repeated approximation}
\begin{theorem}[Composition under changing, certified update rules]
\label{thm:recursive-error}
Let $x_{t+1}=T_t(x_t)$ be a reference process and
$y_{t+1}=\widehat T_t(y_t)$ an approximate process. Assume $T_t$ is
$\kappa_t$-Lipschitz on a region containing $x_t,y_t$, and a certificate at
$y_t$ verifies
$\norm{\widehat T_t(y_t)-T_t(y_t)}\leq\epsilon_t$.
Then, pathwise,
\begin{equation}\label{eq:recursive-error}
\norm{y_T-x_T}\leq
\norm{y_0-x_0}\prod_{t=0}^{T-1}\kappa_t
+\sum_{s=0}^{T-1}\epsilon_s
\prod_{t=s+1}^{T-1}\kappa_t.
\end{equation}
In particular, if $\kappa_t\leq\kappa<1$ and
$\epsilon_t\leq\epsilon$, the long-run error is bounded by
$\epsilon/(1-\kappa)$ when the initial error vanishes.
\end{theorem}
\begin{proof}
The triangle inequality gives
\[
\norm{y_{t+1}-x_{t+1}}
\leq\norm{\widehat T_t(y_t)-T_t(y_t)}
+\norm{T_t(y_t)-T_t(x_t)}
\leq\epsilon_t+\kappa_t\norm{y_t-x_t}.
\]
Iterating this scalar recurrence proves \eqref{eq:recursive-error}; summing
the geometric series proves the final assertion.
\end{proof}
No independence or unbiasedness is required. Rules and budgets may be chosen
adaptively from past computation, provided the displayed Lipschitz and local
error claims remain valid. For an adaptively chosen rule sequence, the
reference is the exact process using that same realized sequence. It is not
automatically the counterfactual trajectory of another policy that would
have chosen different rules after seeing different states.
The contraction assumption is substantive. For $T_t(x)=\kappa x$ and
$\widehat T_t(x)=\kappa x+\epsilon$ on the real line, the error is exactly
$\epsilon\sum_{j=0}^{T-1}\kappa^j$. It grows linearly at $\kappa=1$ and
exponentially for $\kappa>1$. A good one-step compression bound alone cannot
establish indefinite stability of a recursive improvement process.
\subsection{Depth does not degrade exact moment summaries}
Suppose data blocks are repeatedly merged in any tree. At each node retain
their total mass, full first-moment sum, and full second-moment sum; also retain
weighted original exemplars with the same mean and no larger second moment.
Merge the exemplars with weights proportional to block mass and apply
variance-nonincreasing affine elimination again. Induction on the tree shows
that the root mean equals the mean of all original observations, while its
exemplar variance is no greater than the full variance. This conclusion has
no factor depending on tree depth. Final equal-weight rounding is performed
against the original moment certificate.
The distinction between exact weighted memory and rounded student output is
essential. Feeding a biased equal-weight output back as if it were the full
original distribution discards the original mean error; repeated compression
then requires Theorem~\ref{thm:recursive-error} or another explicit error
ledger. Exact preservation applies only to the stored features. A later
change of representation need not be reconstructible from old moments. For
example, equally weighted points $-1,1$ and the eight-point multiset
$(-2,2,0,0,0,0,0,0)$ both have mean zero and variance one, but their fourth
moments are one and four. Their first two moments cannot determine the new
feature $x^4$.
These statements give a precise meaning to a certified improvement component:
the system may change its selection rule, budget, and update policy while
preserving explicit empirical objectives and quantified error conditions.
They do not prove that an unrestricted AI improves its own architecture,
capability, or computational efficiency indefinitely. End-to-end speed
depends on the cost of producing and refreshing the certificates.
\section{An explicit transfer to unseen-data mean risk}
\label{sec:population-transfer}
The selection theorem is conditional on observed data. A population statement
requires assumptions about how those data were generated. The following
result supplies one such bridge without assuming that selected observations
remain independent.
\begin{theorem}[Expected population excess for an arbitrary certified selector]
\label{thm:population-transfer}
Let $X_1,\ldots,X_N$ be independent identically distributed vectors with mean
$\mu_P$ and finite variance
$\sigma^2=\E\norm{X-\mu_P}^2$. Let
$\overline X=N^{-1}\sum_iX_i$ and
$V_N=N^{-1}\sum_i\norm{X_i-\overline X}^2$.
Suppose the selected mean $h$ obeys
$\norm{h-\overline X}^2\leq cV_N$ for every realized dataset, with a fixed
constant $c$. Then
\begin{equation}\label{eq:population-transfer}
\E\norm{h-\mu_P}^2
\leq\sigma^2
\left(\sqrt{c\frac{N-1}{N}}+\frac1{\sqrt N}\right)^2.
\end{equation}
The same conclusion holds if the pointwise inequality is replaced by its
conditional expectation over selector randomization.
\end{theorem}
\begin{proof}
Independence and centering give
$\E\norm{\overline X-\mu_P}^2=\sigma^2/N$.
The exact identity
\[
V_N=\frac1N\sum_i\norm{X_i-\mu_P}^2
-\norm{\overline X-\mu_P}^2
\]
therefore yields $\E V_N=(N-1)\sigma^2/N$. Minkowski's inequality in the
space of square-integrable random vectors gives
\[
\bigl(\E\norm{h-\mu_P}^2\bigr)^{1/2}
\leq \bigl(\E\norm{h-\overline X}^2\bigr)^{1/2}
+\bigl(\E\norm{\overline X-\mu_P}^2\bigr)^{1/2}.
\]
Substituting the two bounds and squaring proves the result. The argument uses
only the integrated selection error, so the conditional-expectation version
follows as well.
\end{proof}
For squared population loss, the risk of $h$ equals
$\sigma^2+\norm{h-\mu_P}^2$, so \eqref{eq:population-transfer} bounds its
expected excess over the population mean. The cross term is retained because
the selector can be biased and dependent on the empirical mean. Dropping it
without an additional orthogonality or unbiasedness assumption would be
unjustified. A selector that is conditionally unbiased does eliminate this
cross term, but then Theorem~\ref{thm:unbiased-barrier} prohibits a universal
selection factor smaller than $1/n$.
% Bibliography entries for the root manuscript:
% \bibitem{craig2020} Baharan Mirzasoleiman, Jeff Bilmes and Jure Leskovec.
% Coresets for Data-efficient Training of Machine Learning Models.
% ICML 2020, PMLR 119:6950--6960.
% https://proceedings.mlr.press/v119/mirzasoleiman20a.html
% \bibitem{maalouf2019} Alaa Maalouf, Ibrahim Jubran and Dan Feldman.
% Fast and Accurate Least-Mean-Squares Solvers. NeurIPS 2019.
% https://arxiv.org/abs/1906.04705
% \bibitem{karimi2016} Hamed Karimi, Julie Nutini and Mark Schmidt.
% Linear Convergence of Gradient and Proximal-Gradient Methods Under the
% Polyak--Lojasiewicz Condition. ECML PKDD 2016.
% https://arxiv.org/abs/1608.04636
|