File size: 16,807 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 | \section{The problem and the new mathematical structure}
\label{sec:overview}
Question 1 of Hanneke, Moran, Shlimovich and Yehudayoff's COLT open-problem
note asks for the exact worst loss when a mean learner is trained on selected
original examples \cite{question}. The learner is fixed; the selector controls
its input. The accompanying work establishes the one-dimensional law and a
general sampling bound \cite{erms}. Dewasurendra determines the full
two-example column \cite{pair}. The supplied planar manuscript proves the
all-budget law in affine dimension at most two. We retain its exact argument
and develop extensions whose hypotheses are explicit.
For a nonempty finite multiset $D=(z_1,\ldots,z_N)\subset\R^d$, write
\[
\mu=\frac1N\sum_i z_i,\qquad
V=\frac1N\sum_i\norm{z_i-\mu}^2,\qquad
L_D(h)=\frac1N\sum_i\norm{h-z_i}^2.
\]
The student receives exactly $n$ original examples, with repetition permitted,
and returns their equal-weight mean. Thus
\begin{equation}\label{eq:profile-definition}
F(d,n)=\sup_{D:V>0}\min_{i_1,\ldots,i_n}
\frac{L_D(n^{-1}\sum_{j=1}^n z_{i_j})}{V}.
\end{equation}
Constant datasets receive ratio one; set $F(0,n)=1$ by convention.
A finite multiset may be replicated to
ensure $N\geq n$ without changing its empirical distribution or possible
selections. Consequently that size convention does not change the supremum.
Expanding a square gives the identity
\begin{equation}\label{eq:risk-decomposition}
L_D(h)=V+\norm{h-\mu}^2.
\end{equation}
The excess factor $F(d,n)-1$ is therefore a relative mean-approximation error.
Affine rank, rather than the number of ambient coordinates, will govern the
constructive upper bounds. Arbitrary finitely supported probability weights
are useful inside the teacher, even though every student output remains an
integer count vector divided by $n$.
\subsection{What changes beyond the planar manuscript}
The main additions have different mathematical roles.
\begin{enumerate}
\item An exact analysis of a frequent anchor with rare simplex vertices
gives a new all-budget obstruction family. Its four-dimensional branch is
strictly above the planar law for every $n\geq2$; for fixed dimensions
greater than four it changes the leading asymptotic lower coefficient.
\item A pair-seeded constructive profile covers every affine dimension and
every budget. A fractional-residual construction supplies sharper
instance-dependent certificates and an $O(n^{-2})$ bound for fixed geometry.
\item An exact finite minimax formulation accepts arbitrary feasible integer
count sets. It treats caps, quotas and cost restrictions without assuming
that an unrestricted theorem survives added constraints.
\item A scalar selection theorem becomes a matrix policy, a feasibility
guarantee for linear margins, and a quadratic-ERM perturbation theorem.
Gradient reuse and recursive update rules then receive explicit error laws.
\end{enumerate}
The lower surface $B_{d,n}$ and the universal upper surface $U_d(n)$ are not
generally equal. The distinction is substantive: a new obstruction is a
proved advance without constituting a complete formula for the entire open
problem. All claims in this manuscript can be read without the predecessor.
\subsection{Notation and quantifiers}
For probabilities $p_i>0$ on points $x_i$ in a finite-dimensional real Hilbert
space, let
\[
\mu_p=\sum_i p_i x_i,\qquad
V_p=\sum_i p_i\norm{x_i-\mu_p}^2,
\qquad \Sigma_p^x=\sum_i p_i(x_i-\mu_p)(x_i-\mu_p)^\top.
\]
Zero masses are discarded. The symbol $\Sigma_p=\diag(p)-pp^\top$ later
denotes covariance in the space of atomic counts; it is distinguished from
the covariance $\Sigma_p^x$ of geometric observations. A positive semidefinite
metric is fixed before selecting unless a randomized-policy theorem explicitly
states a stronger common-policy guarantee. A statement about the best
selection is an existence statement; the supplied unrestricted algorithm is
required to meet its upper bound, not to optimize each individual dataset.
\section{Small weighted memory and exact base certificates}
\label{sec:compression}
\begin{lemma}[Variance-nonincreasing affine compression]
\label{lem:compression-general}
For any finitely supported distribution of affine rank $r$, there is a
distribution on at most $r+1$ of its original points having the same mean
and no greater variance.
\end{lemma}
\begin{proof}
If more than $r+1$ masses are positive, the augmented columns $(1,x_i)$ are
linearly dependent. Choose a nonzero vector $a$ with
$\sum_i a_i=0$ and $\sum_i a_ix_i=0$. Reverse its sign if necessary so that
$\sum_i a_i\norm{x_i}^2\leq0$. Because $a$ has both signs, the positive step
\[
t=\min_{a_i<0}\frac{p_i}{-a_i}
\]
is defined. The new weights $p_i+ta_i$ are nonnegative, have sum one, preserve
the mean, and do not increase the second moment. At least one positive mass
vanishes. Iterate. Since the mean is unchanged, subtracting its squared norm
from both second moments gives the variance comparison.
\end{proof}
This classical affine argument preserves the correct inequality direction.
It does not provide unweighted memory and does not preserve every possible
future feature. Its output weights belong to the teacher's internal
representation; they are not passed to the equal-weight student.
\begin{proposition}[Streaming and mergeable realization]
\label{prop:streaming}
The preceding reduction can be maintained in one pass, with at most $r+1$
resident original observations after each update and at most $r+2$ before
elimination. Keep also the prefix count, first-moment sum and second-moment
sum. After every update, resident weights reproduce the full prefix mean
and have no larger second moment.
\end{proposition}
\begin{proof}
After $t$ observations, scale resident weights by $t/(t+1)$ and append the
new observation with weight $1/(t+1)$. Its mean equals the $(t+1)$-prefix
mean. Its second moment is at most the full prefix second moment by the
induction hypothesis. Apply Lemma~\ref{lem:compression-general}. The same
argument merges weighted block summaries using their original masses.
\end{proof}
An implementation may eliminate any affine dependence, even when fewer
than $r+2$ points are resident. It then ends with an affinely independent
support. The number of exact arithmetic registers is bounded; their bit
lengths may grow with the data and stream length.
\begin{lemma}[The sharp two-example atomic certificate, prior result]
\label{lem:pair}
For a distribution on at most $m\geq2$ points in a real Hilbert space, some
two-point average has squared error at most
\begin{equation}\label{eq:pair-constant}
a_{m-1} V_p,\qquad
a_r=\max\left\{\frac13,\frac{r-1}{2r}\right\}.
\end{equation}
\end{lemma}
\begin{proof}
We rederive the certificate of \cite{pair}. Center the points. If
$p_j\geq1/4$, assign the pair $(j,j)$ mass $(4p_j-1)/3$ and each $(j,i)$,
$i\ne j$, mass $4p_i/3$. Their expected squared mean equals $V_p/3$ by
$\sum p_i x_i=0$.
Otherwise put
\[
S=\sum_i\frac{p_i}{1-4p_i},\quad
C=\frac{S+1}{3S+1},\quad
b_i=\frac{Cp_i}{(1+S)(1-4p_i)},\quad
w_{ij}=4(p_ib_j+p_jb_i)\quad(i<j).
\]
Direct summation gives $\sum_{i<j}w_{ij}=1$. For the pair-frequency vector
$q=(e_i+e_j)/2$, its second moment satisfies
\[
\E qq^\top=C\diag(p)+pb^\top+bp^\top.
\]
Multiplication by the centered support Gram matrix annihilates the last two
terms; expected squared error is $CV_p$. Convexity of
$t/(1-4t)$ gives $S\geq m/(m-4)$, hence
$C\leq(m-2)/(2(m-1))$. A realization has error no greater than its expectation.
\end{proof}
For completeness, sharpness uses two elementary distributions. A mass $1/4$
at one and $3/4$ at zero gives excess $1/3$. Uniform unit vertices of a regular
$(m-1)$-simplex give excess $(m-2)/(2(m-1))$ for every distinct pair, and
repeated pairs cost one. These establish the maximum in
\eqref{eq:pair-constant}. The pair theorem is an input credited to prior work,
not a new result of this extension.
\section{All-budget upper bounds in arbitrary dimension}
\label{sec:upper}
\begin{lemma}[Exact one-unit removal]
\label{lem:removal}
Suppose $n\geq2$ and $p_j\geq1/n$. Define
\[
q_i=\frac{np_i-\mathbf1\{i=j\}}{n-1},\qquad
\nu=\mu_q,\qquad W=V_q.
\]
Then
\begin{align}
\nu&=\frac{n\mu_p-x_j}{n-1},\notag\\
W&=\frac n{n-1}V_p-
\frac n{(n-1)^2}\norm{x_j-\mu_p}^2.\label{eq:removal-variance}
\end{align}
If $b$ averages $n-1$ support points and $h=(x_j+(n-1)b)/n$, then
\begin{equation}\label{eq:removal-error}
h-\mu_p=\frac{n-1}{n}(b-\nu).
\end{equation}
\end{lemma}
\begin{proof}
The masses are nonnegative and sum to one. The mean identity is immediate.
Translate $\mu_p$ to zero. The new raw second moment is
$(nV_p-\norm{x_j}^2)/(n-1)$ and its mean is $-x_j/(n-1)$. Subtract its
squared norm to obtain \eqref{eq:removal-variance}; substitute the mean
identity to obtain \eqref{eq:removal-error}.
\end{proof}
\begin{theorem}[Propagation from any admissible base budget]
\label{thm:propagation}
Assume a constant $c_b>0$ is valid for every distribution on at most $m$
points at budget $b\geq m-1$. Then for every $n\geq b$ the constant
\begin{equation}\label{eq:propagated}
c_n=\frac{bc_b}{n(1+c_b)-bc_b}
\end{equation}
is valid for the same support class.
\end{theorem}
\begin{proof}
Induct on $n$. At the next budget, if some point has squared distance at
most $c_nV_p$, repeat it $n$ times. Otherwise all squared distances exceed
$c_nV_p$. Since $n\geq m$, a largest mass is at least $1/n$. Remove one unit
from that mass and use the $(n-1)$-budget result. The resulting squared
error is at most
\[
\frac{c_{n-1}}n\bigl((n-1)V_p-\norm{x_j-\mu_p}^2\bigr)
\leq\frac{c_{n-1}}n(n-1-c_n)V_p.
\]
This equals $c_nV_p$ when
$c_n=(n-1)c_{n-1}/(n+c_{n-1})$. Writing $a_n=1/c_n$ gives
\[
\frac{a_n}{n}-\frac{a_{n-1}}{n-1}=\frac1{n(n-1)}.
\]
Telescoping yields $a_n=n(a_b+1)/b-1$ and proves
\eqref{eq:propagated}. The zero-variance case is exact.
\end{proof}
\begin{corollary}[Exact planar and one-dimensional profiles]
\label{cor:planar}
For $r=1,2$ and every $n\geq1$,
\[
F(r,n)=1+\frac1{2n-1}=\frac{2n}{2n-1}.
\]
The same guarantee holds for data of affine rank at most two in any ambient
Hilbert space, using at most three resident original observations.
\end{corollary}
\begin{proof}
Some support point has squared distance at most $V_p$. For at most two
points use Theorem~\ref{thm:propagation} with $b=1,c_b=1$; for at most three
use $b=2,c_b=1/3$ and Lemma~\ref{lem:pair}. Compress by affine rank first.
For a matching finite lower bound take $2n-1$ zeros and one unit point. Its
mean is $1/(2n)$, variance $(2n-1)/(4n^2)$ and closest $n$-point grid means
have squared error $1/(4n^2)$. Their excess ratio is $1/(2n-1)$.
\end{proof}
\begin{lemma}[Padding a certified selection]
\label{lem:padding}
If $c_b$ is valid at budget $b$, then for $n\geq b$ the constant
\begin{equation}\label{eq:padding}
\frac{b^2c_b+n-b}{n^2}
\end{equation}
is valid without a support-size restriction beyond that of the base result.
\end{lemma}
\begin{proof}
Fix a certified $b$-point mean and add $n-b$ independent draws from the
original distribution. Their centered sum has mean zero and expected squared
norm $(n-b)V_p$. The squared error of the full sum therefore has expectation
at most $b^2c_bV_p+(n-b)V_p$. At least one realization attains this value.
It can be found deterministically by choosing each additional point to
minimize the conditional expected final squared error; the minimum is no
greater than the average under the sampling distribution.
\end{proof}
\begin{theorem}[Explicit constructive upper profile]
\label{thm:upper-profile}
Set $U_0(n)=0$, $U_r(1)=1$ for $r\geq1$, and
$U_r(n)=1/(2n-1)$ for $r=1,2$. For $r\geq3$ define
\[
a_r=\max\{1/3,(r-1)/(2r)\},\qquad
b_r=\frac{r-2+4a_r}{r^2},
\]
and, for $n\geq2$, set
\begin{equation}\label{eq:upper-profile}
U_r(n)=
\begin{cases}
\displaystyle\frac{n-2+4a_r}{n^2},&2\leq n\leq r,\\[7pt]
\displaystyle\frac{r b_r}{n(1+b_r)-r b_r},&n\geq r.
\end{cases}
\end{equation}
Every dataset of affine rank at most $r$ has an $n$-point selected mean with
\begin{equation}\label{eq:upper-guarantee}
\norm{h-\mu}^2\leq U_r(n)V,
\qquad F(r,n)\leq1+U_r(n).
\end{equation}
The two branches agree at $n=r$. For $r\geq1,n\geq2$, $U_r(n)<1/n$.
All selected points can come from one weighted support of at most $r+1$
original examples.
\end{theorem}
\begin{proof}
Compress using Lemma~\ref{lem:compression-general}. The pair constant is
$a_r$. Lemma~\ref{lem:padding} with $b=2$ proves the first branch and yields
$b_r$ at budget $r$. Theorem~\ref{thm:propagation} with support $m=r+1$ and
base budget $r$ proves the second branch. The planar and rank-one cases
were established above. Since $a_r<1/2$, the first branch is strictly less
than $1/n$. For the second branch, $b_r<1/r$ implies
\[
n(1+b_r)-r b_r-nr b_r=n-n(r-1)b_r-rb_r>0
\]
for $n\geq r$: its minimum under $b_r\leq1/r$ is $n/r-1$, with strictness
at equality. This is equivalent to $U_r(n)<1/n$.
\end{proof}
This upper profile is explicit and constructive; it is not asserted sharp
when $r\geq3,n\geq3$. For example,
\[
U_3(n)=\frac{21}{34n-21}\quad(n\geq3),\qquad
U_4(n)=\frac{28}{39n-28}\quad(n\geq4).
\]
The sharp pair values remain $U_3(2)=1/3$ and $U_4(2)=3/8$.
The dependence on the chosen support bound is safe even if later peeling
reduces the actual support. The implementation retains the original bound
through each recursive call, avoiding an invalid change of induction class.
\section{Fractional residuals and faster instance-dependent decay}
\label{sec:residual}
\begin{theorem}[Residual allocation certificate]
\label{thm:residual}
Given probabilities $p_i$ on $m$ support points, let
\[
a_i=\lfloor np_i\rfloor,\quad f_i=np_i-a_i,\quad
k=\sum_i f_i=n-\sum_i a_i.
\]
If $k=0$, the mean is represented exactly by counts $a_i$. Otherwise
$1\leq k\leq m-1$. Let $q_i=f_i/k$ on its $s$ positive masses, with mean
$\nu$ and variance $W_q$. There is an $n$-point selection with
\begin{equation}\label{eq:residual-bound}
\norm{h-\mu_p}^2
\leq\frac{k^2}{n^2}U_{s-1}(k)W_q
\leq\frac{kW_q}{n^2}.
\end{equation}
\end{theorem}
\begin{proof}
Each fractional part lies in $[0,1)$, so their integer sum is less than $m$.
Use counts $a_i$ and add a certified $k$-point selection from $q$ with mean
$b$. The identity
\[
\mu_p=\frac1n\sum_i a_i x_i+\frac{k}{n}\nu
\]
implies $h-\mu_p=(k/n)(b-\nu)$. Apply
Theorem~\ref{thm:upper-profile} with support-size bound $s-1$ and use
$U_{s-1}(k)\leq1/k$.
\end{proof}
Taking the better of a direct selection and residual allocation gives
\begin{equation}\label{eq:portfolio-bound}
\norm{h-\mu_p}^2\leq
\min\left\{U_{m-1}(n)V_p,
\frac{k^2}{n^2}U_{s-1}(k)W_q\right\},
\end{equation}
with zero as the second term if $k=0$. The residual branch requires at most
$m-1$ nontrivial selection slots, regardless of the requested $n$; the floor
counts can be returned as integer multiplicities without expanding the
entire replay multiset.
\begin{corollary}[Fixed-data quadratic decay]
\label{cor:fixed-data}
For a fixed support of diameter $\mathcal D$, the residual construction gives
\[
\norm{h-\mu_p}^2\leq\frac{(m-1)\mathcal D^2}{2n^2}.
\]
If the distribution and positive variance are fixed, its best normalized
excess is therefore $O(n^{-2})$. If all masses have a common integer
denominator $Q$, every budget divisible by $Q$ permits zero error.
\end{corollary}
\begin{proof}
For any distribution on this support,
$W_q=\tfrac12\sum_{i,j}q_iq_j\norm{x_i-x_j}^2\leq\mathcal D^2/2$.
Substitute in \eqref{eq:residual-bound}. The divisibility statement makes
every $np_i$ integral.
\end{proof}
There is no contradiction with uniform worst-case excess of order $1/n$.
The supremum defining $F(d,n)$ may use a different distribution at every
budget. In the rare-simplex construction below the rare mass, and hence
the variance, decreases as the budget grows. A uniform relative-risk law
cannot be inferred from a fixed-support absolute-error rate.
\begin{proposition}[Uniform-simplex lower bound at arbitrary budget]
\label{prop:uniform-simplex}
For uniform unit vertices of a regular $(m-1)$-simplex and $n=am+s$ with
$0\leq s<m$, the optimal squared error and excess ratio are
\[
\frac{s(m-s)}{(m-1)n^2}.
\]
\end{proposition}
\begin{proof}
For counts $c_i$ summing to $n$, the squared selected mean is
$[m\sum_i c_i^2-n^2]/[(m-1)n^2]$. Moving one unit from a count at least
two larger to a smaller count decreases the sum of squares. Thus the
minimum uses $a+1$ on $s$ vertices and $a$ on the rest. The original
variance is one; substitution proves the formula.
\end{proof}
The uniform-simplex residue law and the rare-line law are both valid lower
bounds. The next section proves that their maximum misses an additional
obstruction, already in dimension four at budget three.
|