Neiru-Siha-II / research /learning_examples.tex
PureOne's picture
Release Neiru Siha II research dataset
19591f3 verified
Raw History Blame Contribute Delete
2.67 kB
\section{A rational quadratic example with a certified reused step}
\label{sec:quadratic-example}
All parts of the quadratic and reuse certificates can be nontrivial on one
exact rational example. Let
\[
u=(3/5,4/5)^\top,\quad
A_\pm=I\pm\tfrac12uu^\top,\quad r_\pm=\pm u,
\qquad
f_\pm(\theta)=\tfrac12\theta^\top A_\pm\theta+r_\pm^\top\theta.
\]
Give the two examples equal population weight. Then $H=I$, $b=0$,
$\theta_*=0$, $M=1/4$ and $G=1$. These probabilities can be represented by a
six-example dataset containing three copies of each example, so the following
budget $n=5$ also satisfies $n\leq N$.
Select three plus examples and two minus examples. The lifted feature cloud
has affine rank one, so $c=1/(2n-1)=1/9$ is available. The selected quantities are
\[
\widehat A=I+\tfrac1{10}uu^\top,\qquad
\widehat r=\tfrac15u,\qquad
\widehat\theta=-\tfrac2{11}u.
\]
The optimized scale in Theorem~\ref{thm:quadratic-erm} is $\tau^2=4/5$.
The exact joint certificate, its universal upper bound, and the actual loss
are respectively
\[
E_{\rm actual}=\tfrac6{125},\qquad
E_{\rm universal}=\tfrac2{15},\qquad
F(\widehat\theta)-F(0)=\tfrac2{121}.
\]
Lemma~\ref{lem:joint-perturbation} applied to the measured certificate gives
$6/235$, while the optimized universal guarantee is $2/25$. Thus
\[
0<\frac2{121}<\frac6{235}<\frac2{25}.
\]
Enumeration of the six possible aggregate count choices independently verifies
that the displayed student also minimizes the original quadratic excess for
this particular dataset and budget. Instance optimality is a property of this
example, not a requirement of the universal selector.
The same selected examples permit a nontrivial reused optimization step. Anchor
at $\theta_0=0$. The exact anchor gradient error is $1/5$, and the Hessian
discrepancy has operator norm $1/10$. Throughout the unit ball, the gradient
error is therefore at most $3/10$. At $\theta=u$,
\[
\nabla F(u)=u,\qquad \widehat g(u)=\tfrac{13}{10}u,
\qquad \epsilon=\tfrac3{10}.
\]
The full objective has smoothness constant $L=1$. The observable rule of
Theorem~\ref{thm:observable-descent} chooses $a=3/13$ and $\eta=10/13$, giving
$\theta^+=u-(10/13)(13/10)u=0$. Both the actual decrease and the certified
decrease are exactly $1/2$. The segment stays inside the unit ball.
The file \texttt{verify\_learning.py} verifies these calculations using exact
rational arithmetic. It also checks an equality case of the joint perturbation
lemma, finite unbiasedness obstructions, the simplex distinction between
deterministic and lottery matrix bounds, and equality in the recursive-error
formula for contracting, neutral, and expanding update rules.