Sequences and Convergence, Pointwise and Uniform Convergence


When does a student or surrogate model match its target everywhere?

Vishnu Boddeti

Example 1

Distillation trains a smaller model to imitate a teacher

When does a distilled student match its teacher everywhere, not just prompt by prompt?

target f narrow net fₙ gap near kink
  • As student capacity, data, or approximation rank grows, its output function may converge for every fixed prompt or input while a moving worst-case region remains poorly matched.
  • Pointwise convergence controls one input at a time; uniform convergence controls the worst input simultaneously.

Pointwise vs. uniform convergence

worst-case gap pointwise gap shrinks
  • A sequence of functions $f_n \to f$ pointwise on a domain if for every fixed $x$, $f_n(x) \to f(x)$ as $n \to \infty$.
  • It converges uniformly if the worst-case gap $\sup_x |f_n(x) - f(x)|$ itself shrinks to $0$.
  • Pointwise convergence can hold everywhere while the max gap refuses to shrink near a kink or jump
  • the same intuition behind the Gibbs phenomenon in Fourier partial sums, and behind why "training loss $\to 0$ pointwise per example" is a weaker claim than uniform control over all inputs.

Increase the sine terms and track the worst error

target $f(x)$ partial sum $f_n(x)$ max pointwise error band
Check the domain edge. Pointwise improvement there does not force the worst error to approach zero.

Example 2

Evaluation convergence is a claim about uncertainty

What does it mean to say an evaluation metric has converged?

N? stays inside band
  • A validation-loss or capability curve can look flat while still producing later spikes.
  • Saying it converges is a precise claim about the entire tail of the checkpoint sequence.
  • What exactly must remain inside a chosen tolerance?

The $\epsilon$–$N$ definition of convergence

L N for all n > N: inside band
  • A sequence $x_n \to L$ means: for every $\epsilon > 0$, there exists an $N$ such that for all $n > N$, $|x_n - L| < \epsilon$.
  • In words — no matter how small a tolerance band you draw around $L$, the sequence eventually enters that band and never leaves it.
  • Shrinking $\epsilon$ generally pushes $N$ further out, but for every $\epsilon$ some finite $N$ must exist.

Shrink $\epsilon$ and find when the sequence stays inside

$\pm\epsilon$ band around $L$ $x_n$ before $N$ $x_n$ from $N$ onward
A smaller tolerance usually requires a larger $N$, but convergence requires some finite $N$ for every positive $\epsilon$.

Example 3

Loss can stabilize before behavior does

Why can validation loss look converged while generated behavior still changes?

  • A scalar metric may form a convergent sequence even when individual prompts continue to change.
  • Reliable model comparison needs a function-level notion of convergence across the evaluation distribution, not only a flat aggregate curve.

Advance training and compare scalar convergence with behavioral stability

Move the control to compare the two normalized responses.
A small change in average loss can coexist with meaningful changes on rare or difficult prompts.

Example 4

Self-refining agents need a stopping rule

When should an agent stop a self-refinement loop?

  • A reflection or repair loop generates a sequence of candidate answers and evaluation scores.
  • It should stop when improvement is below tolerance, the verifier is satisfied, or a hard cost budget is reached—not merely because the text stopped changing.

Allow more refinement rounds and compare answer quality with latency

Move the control to compare the two normalized responses.
Convergence needs a measurable criterion and a maximum horizon; otherwise an agent can oscillate or refine indefinitely.

Practice problems

Each scenario hides a mathematical structure from this lecture. Identify the structure, justify it, and work through the resulting problem.

Problem 1 Confidence calibration

Does a sharpening classifier converge uniformly?

A confidence post-processor repeatedly sharpens scores by sending $x$ to $x^n$. Every fixed score below one eventually looks negligible, but a global deployment guarantee must cover every possible score at once.

Model evidence

$f_n(x)=x^n$ on $[0,1]$.

  1. What function is obtained in the pointwise limit?
  2. Is the convergence uniform? Justify your answer with a supremum or continuity argument.
  3. Why does the distinction matter for a guarantee that must hold across all inputs?
Problem 2 Classifier stability

Can a tiny input change flip the prediction rule?

One prediction module returns a smooth probability; another snaps immediately to the nearest hard class. Near the decision boundary, can either module promise that tiny input changes cause only tiny output changes?

Model evidence

Compare $p(x)=1/(1+e^{-x})$ with $h(x)=\mathbf 1\{x\ge0\}$ at $x=0$.

  1. How would you state that local promise using $\varepsilon$ and $\delta$?
  2. Which module satisfies it at the decision boundary?
  3. Can you construct a sequence of tiny perturbations that exposes any failure?
Problem 3 Online model correction

Will repeated calibration updates settle uniformly?

Each deployment round adds a smaller correction to the calibration curve. At any fixed confidence the correction fades, but monitoring cares about the worst confidence value still in service.

Model evidence

$g_n(x)=\frac{x}{1+nx}$ on $[0,1]$.

  1. What is the pointwise limit of the corrections?
  2. Where is the largest correction, and does it vanish uniformly?
  3. How many rounds guarantee that every correction is smaller than a chosen tolerance $\varepsilon$?