What does The Limits of Guardrail Compression conclude, in numbers, about “Opposing Evidence: Finite Samples, Learnt Compression and How Discriminative Power Is Scored”?

Answer: Chapter 6 sets out the opposing evidence: with 50 drafts, 0.920 of repeated runs choose the wrong knee, and six rules learnt from few records keep only 0.017 bits.

Scope of the numbers: Chapter 1 builds an illustrative source of eight kinds of draft and three verdicts and the counting bound: three rules keep 0.960 of the verdict information, two only 0.629. Chapter 4 shows that the critical exchange rate is the reciprocal of the squared maximal correlation, 0.6844 and 1.4610 on the illustrative source while information rises continuously as rules split.

中文版

Data table

The numbers below are printed verbatim by the chapter's Python program and can be reproduced by running it. standard-library Python onlyprinted values that match the answer: 0.920, 50

true I(X;Y) = 0.7534 bits, true knee = 3 rules
nplug-inbias10.1/nMiller-Madowknee!=3mean knee
500.9075+0.15410.20200.70550.9204.28
1000.8366+0.08320.10100.73560.9204.20
2000.8083+0.05490.05050.75780.8103.96
4000.7786+0.02520.02520.75330.7803.81
16000.7600+0.00670.00630.75370.2803.28
Show the full program
# Chapter 6. p61: the curve you measure is not the curve that exists.
# Draw n labelled drafts from the guardrail source of p11, estimate
# I(X;Y) by plug-in, and pick the knee (fewest agglomerative rules that
# keep 95 percent of the estimated information). Plug-in information is
# biased upwards by about (|X| - 1)(|Y| - 1) / (2 n ln 2) bits (Miller-
# Madow; Paninski 2003); the knee moves with it. 200 seeded repeats.
import math
import random

PX = [0.40, 0.18, 0.10, 0.07, 0.10, 0.06, 0.05, 0.04]
PYX = [[0.92, 0.06, 0.02], [0.25, 0.70, 0.05], [0.30, 0.55, 0.15],
       [0.10, 0.60, 0.30], [0.05, 0.15, 0.80], [0.03, 0.12, 0.85],
       [0.02, 0.08, 0.90], [0.01, 0.04, 0.95]]
CELLS = [(x, y) for x in range(8) for y in range(3)]
WEIGHTS = [PX[x] * PYX[x][y] for x, y in CELLS]


def mutual(groups):
    tot = math.fsum(math.fsum(g) for g in groups)
    py = [math.fsum(g[y] for g in groups) / tot for y in range(3)]
    terms = []
    for g in groups:
        pg = math.fsum(g) / tot
        for y in range(3):
            v = g[y] / tot
            if v > 0:
                terms.append(v * math.log2(v / (pg * py[y])))
    return math.fsum(terms)


def knee(groups, share=0.95):
    full = mutual(groups)
    groups = [list(g) for g in groups if math.fsum(g) > 0]
    best_k = len(groups)
    while len(groups) > 1:
        cand = None
        for i in range(len(groups)):
            for j in range(i + 1, len(groups)):
                m = [a + b for a, b in zip(groups[i], groups[j])]
                rest = [g for k, g in enumerate(groups) if k not in (i, j)]
                v = mutual(rest + [m])
                if cand is None or v > cand[0]:
                    cand = (v, rest + [m])
        groups = cand[1]
        if cand[0] >= share * full:
            best_k = len(groups)
    return best_k


true_groups = [[PX[x] * PYX[x][y] for y in range(3)] for x in range(8)]
true_i = mutual(true_groups)
print("true I(X;Y) = %.4f bits, true knee = %d rules"
      % (true_i, knee(true_groups)))
rng = random.Random(20261113)
print("     n   plug-in  bias    10.1/n  Miller-Madow  knee!=3  mean knee")
for n in [50, 100, 200, 400, 1600]:
    est, mm, wrong, ks = [], [], 0, []
    for _ in range(200):
        counts = [[0.0] * 3 for _ in range(8)]
        for x, y in rng.choices(CELLS, weights=WEIGHTS, k=n):
            counts[x][y] += 1.0
        v = mutual(counts)
        est.append(v)
        mm.append(v - 7.0 / (n * math.log(2)))
        k = knee(counts)
        ks.append(float(k))
        if k != 3:
            wrong += 1
    mean = math.fsum(est) / 200
    print("  %4d   %6.4f  %+6.4f  %6.4f   %6.4f        %5.3f    %4.2f"
          % (n, mean, mean - true_i, 7.0 / (n * math.log(2)),
             math.fsum(mm) / 200, wrong / 200, math.fsum(ks) / 200))

Single source and citation

Fan, Y. (2026). The Information Bottleneck and Rate-Distortion Limits: How Far a Bilingual Monograph Can Be Compressed into Guardrails before Discriminative Power Falls Away (Version 1.0.0), Chapter 6: Opposing Evidence: Finite Samples, Learnt Compression and How Discriminative Power Is Scored. 5D Library. DOI pending

DOI pending · monograph page goes live after publication · 5DL-B-250 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026250,
  author    = {Fan, Yafu},
  title     = {The Information Bottleneck and Rate-Distortion Limits: How Far a Bilingual Monograph Can Be Compressed into Guardrails before Discriminative Power Falls Away},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

Other chapter answers in this book