Notes 28: mathematical corrections and scope

Source snapshot: source/frequency_est.tex.
Original: /home/sariel/rand_alg/notes/28_frequence_est/frequency_est.tex.
Existing review consulted: source/review.md. The original notes were not edited.

1. Averaging uses at least one sample, finite variance, positive mean, and
   independent samples. Chebyshev is applied with a positive absolute
   threshold, including variance-zero cases.
2. The number of blocks is an odd positive integer; total samples are
   alpha times beta. Median failure implies a bad majority, not conversely.
   Logarithmic bounds use log(2/delta) for the whole range 0 < delta < 1.
3. Reservoir replacement resets the suffix counter even for the same key.
   All frequency sums use universe size n; stream length is m.
4. The general estimator's second moment is bounded above, not equal to
   km F_(2k-1). The variance proof uses F_(2k-1), not F_(2k-2).
5. Parameters use the proved upper bound k n^(1-1/k) on relative variance,
   not the unknown target mean/variance. Arithmetic/word bounds are stated
   for fixed integer k. Empty streams return zero.
6. The polynomial hash produces ternary values. Keep zeros and normalize
   T^2 by q=(p-1)/p. Never condition on all outputs being nonzero: such
   conditioning need not preserve independence. The repaired variance is
   2 F_2^2 + (1/q-3) F_4 <= 2 F_2^2. An odd prime p>n of size O(max(n,2))
   suffices, with independent degree-at-most-five coefficient vectors.
7. The ideal-sign calculation remains as the pedagogical first step; the
   finite-seed theorem uses the normalized ternary construction.
8. F_2 amplification uses an explicit exponential-moment bound and rounded,
   odd block counts. The stated update cost assumes unit-cost word arithmetic;
   prime selection is preprocessing.

Content follows Chapter 28 and its imported probability/hash prerequisites.
Diagrams instantiate the stated definitions; no additional streaming methods
are introduced. Full-deck timing: 5 + 15 + 25 + 25 + 5 = 75 minutes.